OIDS-45: A large-scale benchmark insect dataset for orchard pest monitoring

doi:10.21203/rs.3.rs-4339725/v1

Download PDF

Article

OIDS-45: A large-scale benchmark insect dataset for orchard pest monitoring

https://doi.org/10.21203/rs.3.rs-4339725/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Insects play a crucial role in agricultural production and should not be overlooked. However, there is currently no large-scale dataset available specifically for common insects in orchards. Additionally, datasets for computer vision target detection tasks are limited in the field of insects, which hinders the use of deep learning target detection techniques in orchard insect monitoring. This paper presents the OIDS-45 dataset, which is a large-scale dataset for orchard insect monitoring. The dataset contains 58,585 images of 45 categories of common insects found in orchards. The dataset exhibits a long-tailed distribution, and all images are labeled with borders, making them useful for target detection tasks. The dataset represents the category of orchard insects and has a larger sample size, more categories, and more features in the orchard scenario than previous datasets. We compared our dataset with existing typical insect datasets using advanced target detection algorithms to evaluate its features and quality. The experimental results indicate that current target detection algorithms are not yet capable of accurately identifying and detecting insects in orchards. This is due to the small size of individual insects, the morphological similarities between some species, and the existence of multiple growth stages in some insects. The production and release of this dataset aim to support research in the fields of orchard pest control and insect monitoring in orchards.

Biological sciences/Computational biology and bioinformatics/Data acquisition

Biological sciences/Computational biology and bioinformatics/Image processing

Biological sciences/Computational biology and bioinformatics/Machine learning

deep learning

object detection

insect dataset

insect detection

It is widely acknowledged that the insect situation in orchard production has a significant impact on fruit crop productivity, both in terms of economic losses caused by pests and in terms of rationalizing the use of beneficial insect populations for biological pest management [1]. Insect classification and identification are also critical in orchard insect monitoring, which is significant for the development of the fruit sector as well as the agricultural economy's security and stability [2]. Today, the orchard fruit planting sector is massive. People's living conditions are gradually improving as social and economic growth accelerates, the public places an increasing value on health, and demand for fruits continues to rise. In China, for example, fruit production in 2022 was 312,962,400 tons, a 4.42% increase over the previous year [3]. At the same time, as fruit production increases, the losses caused by pests and illnesses become more apparent. Orchard pests lead to fruit yield reductions and seriously lead to the destruction of fruit trees. According to statistics, at present, the loss of China's orchard yield due to pests and diseases accounts for about one-tenth of the total cost, and the cost of controlling orchard pests and diseases accounts for about two-fifths of the total cost [4]. In Shaanxi Province, China, for example, in 2023, the major orchard pests and diseases had a moderate occurrence, affecting an area of more than 32 million hectares, mainly caused by pear louse, fruit fly, and mesquite, and causing direct economic losses of up to 700 million yuan.

However, identifying insects can be challenging due to their small size, wide distribution, and complexity. Therefore, it often requires the expertise of biologists to identify them manually using the naked eye. This process can be time-consuming and costly, which limits the accuracy and efficiency of the analysis. The automatic monitoring of insect conditions in orchards has gained increasing attention due to the booming development of deep learning and computer vision technology, particularly target detection technology, in recent years.

Large-scale insect databases based on orchard conditions are scarce. Wu et al. provided a representative dataset, IP102 [5], which includes 102 common insect species. However, only a small percentage of these species are found in orchards, and they are primarily common in paddy fields and field crops. In addition to IP102, most available insect datasets for orchard settings lack the necessary quantity and quality for research and applications.

Additionally, many insect datasets tend to prioritize the identification of pests while overlooking the significance of beneficial insects. In agriculture, beneficial insects can act as biological control agents in organic farming [6], thereby reducing the need for chemical pesticides and fertilizers. Ecological farming [7] widely employs beneficial insects to regulate pest populations and minimize crop damage. Therefore, our proposed dataset will not only focus on the study of pests but also emphasize the significance of beneficial insects.

In practice, identifying different insect species can be challenging due to their physical similarities [8]. Achieving accurate identification requires both high levels of expertise and a large amount of image data. Therefore, recognizing these various bug species with high similarities is a highly difficult problem for object detection.

The use of deep learning has led to a significant improvement in object detection algorithms. These algorithms utilize deep learning neural networks for feature extraction and classification, resulting in better performance [9]. Meanwhile, in terms of localization and recognition, object detection approaches outperform classification techniques [10]. They are also suited for multi-target processing [11], are scalable, capable of handling various and complicated scenarios, and meet real-time needs [12]. Therefore, we consider that target detection technology has an advantage over classification technology in orchard insect identification. This is because it can provide more comprehensive information, accurately determining the species of insects and their positions in the image. Target detection technology enables real-time monitoring and identification of insects in the orchard, allowing for timely and targeted control efforts to preserve crops and reduce orchard damage. Its advantage over categorization technology in agricultural spraying lies in its ability to precisely locate and identify pests. Agricultural workers can use target detection technology to accurately identify areas with high pest activity and concentrate pesticide spraying on these regions. This reduces pesticide use, improves application efficiency, and minimizes negative impacts on beneficial insects and the environment. This enhanced management technique, combined with target detection technology, can effectively protect orchard crops and increase agricultural productivity. These characteristics enhance target detection and increase its practical applicability.

To advance the study of insects in orchards and agricultural areas through computer vision target detection techniques, we propose a new large-scale insect dataset dedicated to orchard scenarios. As shown in Fig. 1, the insects in each image in the dataset belong to a specific species. This article will describe in detail the construction of the dataset. Meanwhile, to demonstrate the practicality of the OIDS-45 dataset, we utilized state-of-the-art target detection algorithms to evaluate and compare its performance against other datasets. The results indicate that our dataset presents a challenge.

To summarize, we consider our contributions are as follows:

To the best of our knowledge, there is no large-scale insect dataset specifically designed for orchard scenarios. We have developed a comprehensive insect dataset for orchard applications, which includes both beneficial and pest insects. The entire dataset is open source for academic use.
Extensive experiments were conducted on our dataset and other representative insect datasets from various application scenarios, such as IP102. Several cutting-edge target detection models were used to contribute to the research on insect target detection and recognition in agricultural settings.

2.1. Related work

This section presents existing methods and datasets for insect identification, as well as related work that has been done in this area.

2.1.1 Insect identification technology

Pests can significantly impede agricultural development. Accurate pest identification is essential for predicting and controlling pests. Traditional manual pest identification methods have been historically inefficient, prone to human subjective judgment, and unsuitable for large-scale agricultural production scenarios. Since 2003, researchers have attempted to detect target pests using classic machine learning techniques such as support vector machines (SVM) [13], K-nearest neighbor (KNN) [14], classification algorithms, and AdaBoost [15]. To address this issue, researchers have developed new methods that use unsupervised learning to detect target pests without relying on pre-designed features. However, these strategies have a common flaw: they rely on intentionally pre-designed features. While these techniques have been successful in certain situations, they are not always effective. However, pests in their natural state come in a wide range of forms, sizes, and backgrounds, which can result in low robustness and weak generalization capacity for classic machine learning approaches when identifying them in the real world. Therefore, these methodologies face challenges when it comes to practical implementation and diffusion.

In recent years, there have been significant improvements in insect recognition accuracy and efficiency due to the development of deep learning-based insect recognition techniques. Lu et al. [16] developed an 8-layer CNN [17] network to learn powerful local features from complex insect image datasets and achieved high average accuracy in classifying 12 important rice insect species. Wang et al. [18] utilized LeNet-5 [19] and AlexNet [20] to classify images of crop pests by analyzing the impact of the convolutional kernel and the number of network layers. Yang et al. [21] proposed an insect recognition system based on image saliency analysis and a deep learning model, specifically a convolutional neural network (CNN), which demonstrated good robustness. K. Thenmozhi and U. Srinivasulu Reddy [22] proposed an efficient deep CNN model and performed classification experiments on three publicly available insect datasets. The first dataset used was the Central National Bureau of Agricultural Insect Resources (NBAIR) dataset, which consists of 40 classes of field crop insects with a total of 4500 images. The second and third datasets contained 24 and 40 classes of insects, respectively. However, deep learning-based efforts have been limited by insufficient samples to optimize the large number of hyperparameters in the CNN network, resulting in unstable training processes and poor insect recognition accuracy. To promote the further development and application of pest recognition techniques, it is necessary to address the issue of insufficient insect samples. We created a dataset of orchard insects and labeled them. The dataset includes 45 insect classes and a total of 58,595 samples.

2.1.2 Relevant datasets

Currently, there are limited datasets available on insects. Xie et al. [23] presented a dataset that included 1440 samples from 45 common insect types. The National Bureau of Agricultural Insect Resources (NBAIR) produced a dataset of 4500 photos representing 40 different field crop insect groups. However, each category still has a small number of samples, making training a CNN difficult. Later, IP102 provided an agricultural insect dataset with 75,222 photos depicting 102 prevalent insects. However, out of the approximately 19,000 photos available, only a portion of them include bounding boxes for object detection. In contrast, we have compiled a dataset of 58,595 orchard insect images spanning 45 categories, which exhibit a natural long-tailed distribution. Additionally, all of the images in our dataset provide bounding boxes for object detection, which is crucial for model training.

2.2. Creation of datasets

To construct the dataset, we drew heavily on the advanced concepts and experiences of the IP102 dataset. The specific procedure is as follows:

Establishment of a classification system for common insect species in orchards.
Collecting insect images through multiple means.
Screening and filtering of high-quality images that meet the requirements of the study.
Annotate datasets with professional standard labels.

2.2.1. Preliminary research work

Prior to collecting insect data in orchards, we conducted preliminary research by searching for information and consulting relevant insect experts. We compiled a list of common insects found in orchards, including both pests and beneficial insects, and established a classification system for the OIDS-45 dataset. Our orchard insect dataset is divided into two main categories: pests and beneficial insects. Each category of insect is further classified based on its biological taxonomy using a commonly used three-tier structure. For instance, in an apple orchard, apple aphids are considered pests and are categorized as such.

2.2.2. Image collection

The datasets are mainly obtained from the Internet and through field photography. For instance, the COCO, ImageNet, and IP102 datasets heavily rely on the Internet for image data. To obtain insect image data for each category, we crawled search engines such as Bing, Google, Baidu, or professional agricultural websites. Additionally, we filmed relevant video clips containing insects in the relevant orchards in Sichuan Province, China, capturing the images at a speed of 5 frames per second. This resulted in the collection of almost 170,000 photos as candidates for the OIDS-45 dataset.

2.2.3. Data filtering and selection

Four volunteers participated in manually screening and filtering the image dataset. Entomology experts and professors trained the volunteers on insect morphology, including the different growth cycles of each variety of bug, before the task began. During the data screening and filtering procedure, volunteers removed incorrect images that were blurry, damaged, unrelated to insects, or outside of the designated categories. The screening process divided image quality into three categories: 'OK,' 'Pending,' and 'Sifted Out.' Only images that blatantly fulfilled the specifications were approved. The images were classified as 'to be determined' and sent to entomology specialists for identification if they could not be conclusively deemed correct. Images that were obviously not up to par were immediately filtered out. Additionally, we required information on the resolution of the photos and the size of the insect targets in the target detection task to ensure high-quality images in the dataset. To identify insects under the surveillance camera at a distance of 5 to 15 meters, we ensured that each image's insect size was between 50 and 350 pixels. The OIDS-45 dataset consists of 58595 images of insects, totaling 45 categories.

2.2.4. Data preprocessing

In actual orchard insect monitoring scenarios, such as collecting and recognizing insects with a camera, acquired insect images are often blurred, veiled, or darkened due to environmental factors or being concealed by items such as leaves. During the testing process, we employed suitable methods to preprocess the dataset to enhance its applicability to real-world scenarios, improve the model's generalization capability [24], mitigate the risk of overfitting [25], and enhance the model's resilience [26].

First, we used OpenCV-Python [27] to apply the Gaussian bilateral filtering technique [28] to blur the images of one-tenth of each class of insects. Bilateral filtering is a non-linear method that balances the spatial proximity and pixel-value similarity of an image while considering spatial information and gray-scale similarity to achieve edge-preserving denoising. This technique is simple, non-iterative, and locally processed. Edge-preserving denoising achieves its filtering effect through a filter that consists of two functions: one function determines the filter coefficients based on geometric spatial distance, and the other determines the filter coefficients based on pixel difference. Gaussian bilateral filtering involves three main parameters: filter diameter, color space variance, and gray value variance. The diameter of the filter determines the range of pixels considered during the filtering process. A larger filter will result in a greater blurring effect, while a smaller filter will retain more details. The variance in color space determines the degree of variation between pixels in the color space. A larger color space variance leads to a wider range of color differences being retained, while a smaller color space variance leads to more consistent colors. The gray value variance determines the degree of difference between pixels in gray space. A larger gray-value variance retains more edge information, while a smaller gray-value variance results in a smoother image. The parameters of the Gaussian bilateral filter are set according to the actual needs before performing the filtering process. The Gaussian filter diameter, color space variance, and gray value variance were set to 15, 75, and 75, respectively. These parameter settings will impact the final blurring effect. The principle formula for the bilateral filtering technique is as follows:

$g(i,j)=\frac{\sum _{(k,l)\in s(i,j)}f(k,l)w(i,j,k,l)}{\sum _{(k,l)\in s(i,j,k,l)}w(i,j,k,l)}$

(1)

The equation defines the output points as g(i, j), with S(i, j) representing the range of (2N + 1) sizes centered on (i,j). The multiple input points are represented by f(k, l), and the values computed after two Gaussian functions are represented by w(i, j, k, l).

To replicate the conditions in which insects are occluded under realistic circumstances, we utilized the Cutout [29] occluder data augmentation method to process 10% of each type of insect image. This approach involves obscuring the picture information of the filled region, which aids in enhancing the model's generalization ability. Figure 2 shows that before beginning the cutout processing, it is necessary to establish the parameters, such as the size of the masked region and the fill value. These parameters have a significant impact on the final blurring effect. We set the masking area to 50*50 pixels and the masking fill color to black.

The YOLOV4 paper suggested the Mosaic [30] data augmentation method, which involves cropping four random photographs and splicing them together as training data. The advantage of this approach is that it enhances the image background, and the four images are combined to increase the batch size in a disguised manner [31]. Additionally, the four images are also computed during the batch normalization [32] operation, making it less dependent on its own batch size. As a result, YOLOV4 can be trained on a single GPU [33]. Mosaic data augmentation was utilized for the dataset. Four insect photographs were randomly selected and labeled 1, 2, 3, and 4 for each class. The following operations were performed on each of the four photos:

Determine the location of Image 1, usually choosing to place it in the upper left corner of the larger image.
Determine the location of Image 2, usually choosing to place it in the upper right corner of the larger image.
Determine the location of Image 3, usually choosing to place it in the lower left corner of the larger image.
Determine the location of Image 4, usually choosing to place it in the lower right corner of the larger image.

When selecting the position, ensure that each image maintains consistent size and proportions. Next, combine images 1, 2, 3, and 4 in the designated location to create a larger image. Finally, make any necessary adjustments to the composite image, such as resizing or cropping, to maintain the overall size and proportion. Figure 3 illustrates the schematic arrangement. Mosaic data augmentation is a technique that randomly sets the boundary position and size of each mosaic image (blue box). This technique helps to increase the diversity of the training data and improve the model's robustness. The mosaic image is created by stitching together cropped regions containing segments from four different images. Synthetic training samples are then generated from these mosaic images.

To enhance the training model's robustness and generalizability, we utilized the CosinGAN [34] model for data augmentation and generalization. The CosinGAN model is an improved version of the SinGAN [35] model, which generates more diverse images while preserving the original image structure, thereby reducing overfitting. The CosinGAN data augmentation process utilizes adversarial learning, as shown in Fig. 4. This involves training two networks: the discriminator, which distinguishes between real and generated images, and the generator, which produces images with improved quality and diversity. Additionally, the enhanced image produced by CosinGAN is visually more realistic due to the generator's ability to preserve the input image's structural information using cosine similarity. Finally, CosinGAN learns the mapping relationship between the input image and the target enhancement style, allowing for successful digital image enhancement through continuous optimization of the adversarial process between the generator and the discriminator.

CosinGAN (Contextual Similarity GAN) is an image data improvement method that uses generative adversarial networks (GAN) [36]. It expands the original dataset by learning contextual information from the input image and creating new photos of similar situations. CosinGAN outperforms other generative adversarial network models, such as SinGan, in terms of sample quality, generating resolution, and training speed. In real life, insects can have similar shapes, textures, and structures, making a small dataset insufficient to cover all possible scenarios and variants. For instance, the seven-star ladybird beetle and the heterochromatic ladybird beetle share a morphology and texture that often require professional entomological knowledge to distinguish. Therefore, it is necessary to augment and expand the dataset using methods such as CosinGAN.

Data augmentation operations, such as random angle rotation, brightness adjustment, contrast enhancement, chromaticity sharpness enhancement, and mirror flip [37], were used to increase the quantity of the insect dataset. The dataset was expanded to five times its original size.

2.2.5. Professional data annotation

After the initial screening and filtering of images, the most crucial phase is to label the dataset using entomological expertise. The OIDS-45 collection comprises 45 species of common orchard insects, including both pests and beneficial insects from several families. We consulted relevant biological specialists to explain and identify each type of insect, ensuring that volunteers accurately labeled the data.

During the annotation process of the dataset, four volunteers were invited to annotate the images. Each volunteer was responsible for bounding box annotations for 10 to 12 insect categories using the Labellmg [38] annotation tool. Prior to the annotation process, the volunteers received systematic and professional training on the knowledge related to the corresponding insect species to facilitate correct annotation. The pictures were tagged based on the following criteria:

Accurately labeled target bounding boxes: Ensure that the labeled target bounding box covers the complete area of the target object to avoid omission or mislabelling and to ensure that the model can accurately identify the target.
Avoiding overlapping objectives: In cases where there is overlap between multiple targets, each target needs to be labeled individually to avoid overlapping target bounding boxes that may be difficult for the model to identify.
Handling of small and obscured targets: In addition to labeling the target's bounding box, the target's category information needs to be labeled to ensure that the model is able to correctly classify different categories of targets.
Consistent labeling style: For small and occluded targets, their bounding boxes need to be carefully labeled to ensure that the model can accurately detect and identify these targets.
Avoiding labeling bias: Maintain consistency in annotation style, including the way bounding boxes are drawn, the format of category labels, etc., so that the model can correctly interpret the annotation information.
Handling blurred and complex scenes: In the labeling process, it is important to avoid subjective bias as much as possible, to maintain the objectivity and accuracy of the labeling, and to avoid misleading the model.
Data quality control: For ambiguous or complex scenes, patience in labeling is required to ensure the accuracy and integrity of the bounding box and improve the model's performance in such scenes.

At the end of the labeling process, entomologists were invited to assess each labeled image. They were asked to evaluate whether the categories were accurately labeled, especially for insects with morphological resemblance to others. Three rounds of review were conducted, during which potential errors or inconsistencies were detected and remedied in a timely manner.

2.3. Dataset partitioning

The OIDS-45 dataset comprises 58,595 images and 45 classifications of orchard insects, with an average of around 1,000 samples per category. We segmented the dataset into training, validation, and test sets following an approximate 7:1:2 ratio. Therefore, the OIDS-45 dataset contains 41,016 training images, 5,859 validation images, and 11,720 test images for the target detection task.

Table 1

The insect images of the OIDS-45 dataset were divided according to the ratio of training set, validation set, and test set as 7:1:2. "Super-Class" indicates the number of corresponding orders, and "Class" indicates the number of species under the corresponding order. "Vermin" and "beneficial insects" denote the distribution of beneficial and pest insects in the dataset, respectively.
	Super-Class	Class	Train	Val	Test
Insecta	Hemiptera	6	3441	491	985
	Homotpera	1	1005	143	288
	Coleoptera	15	14965	2137	4277
	Lepidoptera	7	6526	932	1865
	Diptera	5	4972	710	1421
	Orthoptera	4	3425	489	979
	Neuroptera	1	1294	184	371
	Hymenptera	2	1821	260	521
	Araneae	1	879	125	252
	Odonata	1	908	129	261
	Haplotaxina	1	894	127	257
	Mantodea	1	883	126	253
OIDS-45	Vermin	31	27469	3924	7849
OIDS-45	Beneficial insects	14	13547	1935	3871

Table 2

The OIDS-45 dataset labels each class of insects in order from 1–45, and Numbers represents the number of dataset image sheets for each class. Insects marked with an "*" in the species column of Table 2 belong to the category of beneficial insects, while those not marked with an "*" belong to the category of pests.
Number	Species	Numbers	Number	Species	Numbers
1	Cerya purchasi Maskell	1452	24	Cetonia aurata*	1567
2	Viteus vitifoliae	1435	25	Dorcus titanus*	1577
3	Propylea japonica	1462	26	Allomyrina dichotoma*	1398
4	Apple aphis	771	27	Gryllotalpa orintalis	790
5	Cicadella viridis	795	28	Ocanthus longicauda Matsumura*	1490
6	Stephanitis nashi	300	29	Chrysoperla sinica*	1849
7	Aphanostigma piri	149	30	Myzus persicae	1561
8	Potosia brebitarsis	1408	31	Exorita japonica Townsend	1372
9	Buprestoidea*	1448	32	March fly	1335
10	Shining leafchafer	1457	33	Tephritidae	1429
11	Tuberolachnus salignus	1450	34	Ichneumonidae	1463
12	Coccinella septempunctata*	1439	35	Drosophilidae	1410
13	Melolontha serrulate	1453	36	Formicidae	1139
14	Harmonia axyridis*	1446	37	Cydia pomonella	1367
15	Scyphophorus acupunctatus Gyllenhal	1454	38	Eupithecia subtacincta*	1393
16	Agrotis segetum	1290	39	Araneida*	1255
17	Agrilus planipennis	874	40	Aeshna melanictera Selys*	1298
18	Anoplophora chinensis	1321	41	Pheretima*	1278
19	Malacosoma neustria Testacea	1083	42	Locusts	1320
20	Biston betularia	1651	43	Mantodea*	1262
21	Bactrocera caudata	1557	44	Tettigonia viridissima	1294
22	Heliothis virescens	1306	45	Ostrinia furnacalis	1233
23	Epicauta ruficeps*	1514

2.4. Dataset structure

The OIDS-45 dataset is organized hierarchically, as shown in Table 1. For instance, pear aphids, apple tumor aphids, and apple yellow aphids all belong to the Aphididae superclass, which is also a pest category. Table 2 displays the number and numbering of insect images for each category in the dataset. Figure 5 presents the distribution of the number of different subjects and species in the insect dataset in the form of heat maps. The OIDS-45 dataset comprises 14 beneficial and 31 pest species, as previously stated. In Table 2, insects marked with an "*" are classified as beneficial insects, while those without the asterisk are classified as pests.

2.5. Comparison with other datasets

Table 3 shows a comparison between the OIDS-45 dataset and other insect datasets. Our dataset has 45 categories fewer than the classic IP102 dataset, but it is more relevant as it is based on the orchard scenario. Additionally, our dataset has 58595 labeled images, which is significantly more than the 10,000 labeled images in IP102, making it more advantageous and interesting for practical applications in target detection research. Compared to other existing insect datasets, our dataset contains over ten times the number of samples. Additionally, the maximum and minimum categories of other datasets are 40 and 8, respectively. To ensure accurate results, it is necessary to have a large number of images for each class of insects during training, given the similarity in morphology among different insects in real-life situations. Our dataset contains an average of 1000–1300 images per class of insect, which exceeds the average of other datasets. In the target detection scenario, the limited number of samples can cause an unstable training process, resulting in poor generalization ability and overfitting. This is especially true for insect objects with different growth stages that are difficult to distinguish. Furthermore, only half of the available insect datasets are usable, and most of the existing insect-related datasets are not easily applicable to real-world situations, which reduces their research significance.

Therefore, we feel that the OIDS-45 dataset addresses two distinct challenges and shortcomings:

The insect classes in the OIDS-45 dataset are all in orchard contexts. There is no large-scale insect dataset exclusively for orchard scenarios; thus, our dataset proposal will cover that need.
Due to the morphological similarity of insects, this poses a challenge to the task of detecting insects using target detection techniques. A sufficient, diverse, and balanced large-scale dataset can help the target detection model better learn the target features, improve detection accuracy and generalization ability, and thus achieve better target detection results for insects. The OIDS-45 dataset contains an average of 1,000–1,300 images for each class of insect, which is a sufficient amount of data.

Table 3

Comparison with existing insect-related datasets." Class" denotes the number of classes, "Avail" denotes whether the dataset is available, "Sample" denotes the number of samples in the dataset, and "Avg Sample" indicates the number of samples in the dataset, and "Avg" indicates the average number of samples per category.
Dataset	Year	Class	Avail	Sample	Avg
Samanta et al.[39]	2012	8	N	609	76
J. Wang et al.[40]	2012	9	Y	225	25
Venugoban et al.[41]	2014	20	N	200	10
Xie et al.[42]	2015	24	Y	1440	60
Liu et al.[43]	2016	12	N	5136	428
Chengjun Xie et al.[44]	2018	40	N	4500	113
AI Challenger 2018[45]	2018	27	Y	45285	1677
IP102[46]	2019	102	Y	75222	737
Deng et al.[47]	2020	10	Y	563	56
ArTaxOr[48]	2021	7	Y	15376	2197
BAU-Insectv2[49]	2024	9	Y	2089	232
OIDS-45	2024	45	Y	58595	1302

2.6. Evaluation Metrics

Several composite metrics were used to validate our dataset for the target detection task, including precision [50], recall [51], mean average precision (mAP) [52], model size (Params), and model computation (Flops).

Accuracy is a popular assessment statistic for target detection tasks, measuring the ratio of accurately detected targets to the total number of targets detected. The formula for accuracy is as follows:

$$Pre=\frac{TP}{TP+FP}$$

In the formula, TP represents the true positive sample, and FP represents the false positive sample.

The recall of a target detection task is the percentage of correctly detected targets by the model among all real targets. It indicates the ability to find all positive samples in a specific category. This is an important index for evaluating the performance of a target detection model. The formula is as follows:

$$Recall=\frac{TP}{TP+FN}$$

In the formula, TP represents the true positive sample, and FN represents the false negative sample.

IOU (Intersection over Union) is a measure of the degree of overlap between two bounding boxes in the target detection task. The accuracy of the detection results is measured by calculating the ratio of the area of their intersection to the area of their concatenation. The value of IOU ranges from 0 to 1, and the closer the value is to 1, the higher the overlap between the detection results and the real target box. IOU is often used to evaluate the performance of target detection algorithms. Mean accuracy (mAP), mAP.50, and mAP.75 were used as performance evaluation metrics. IoU=[.50:.05:.95] was used to measure the IOU threshold ranging from 0.5 to 0.95, with an increase of 0.05 each time. This setting can evaluate the performance of the target detection model under different overlap requirements and help to understand the robustness and accuracy of the model under different accuracy requirements.

In evaluating the complexity and computing efficiency of a target detection model, it is important to consider metrics such as model size (measured in M, or million parameters) and model computation (measured in Flops). In evaluating the complexity and computing efficiency of a target detection model, it is important to consider metrics such as model size (measured in M, or million parameters) and model computation (measured in Flops). These metrics, respectively, measure the model's storage requirements and computational efficiency. In evaluating the complexity and computing efficiency of a target detection model, it is important to consider metrics such as model size (measured in M, or million parameters) and model computation (measured in Flops). The formula to compute the number of parameters (params) for popular neural network layers, such as convolutional or fully connected layers, is as follows:

$Params={K}_{ℎ}\times {K}_{w}\times$ Cin$\times$Cout(4)

In the equation, ${\text{K}}_{\text{ℎ}}$ and ${\text{K}}_{\text{w}}$ represent the convolution kernel's height and breadth, Cin the number of input channels, and Cout the number of output channels. It is crucial to note that this is only a simple computation; the actual model may contain numerous such layers as well as other parameters (for example, bias terms). To calculate the number of parameters for the entire model, we must add the number of parameters for all of these levels.

FLOPS stands for floating-point operations per second, which is a measure of hardware performance. Lowercase FLOPS (floating-point operations) represent the number of floating-point operations used to assess the complexity of an algorithm or model. The FLOPS of a convolutional layer can be estimated using the following formula:

$$flops=2\times {ℎ}_{2}\times {w}_{2}\times {k}^{2}\times {c}_{1}\times {c}_{2}$$

The height and width of the output feature map are represented by ${ℎ}_{2}$ and ${w}_{2}$, respectively. The convolution kernel's size is represented by k, and the input and output channel counts are represented by ${c}_{1}$ and ${c}_{2}$, respectively. To calculate the flops of the entire model, add the flops of all the layers in the model.

In the deployment of insect detection algorithms in the field, resource constraints such as hardware often exist, which place requirements on model lightweighting and computational efficiency. Algorithm complexity and model size are closely related when deployed on hardware devices. Increased algorithm complexity requires more computational resources to execute the algorithm in real-world scenarios in orchards, which can result in longer execution times and higher energy consumption. Additionally, larger model sizes require more memory and storage space to load and run the model, which may limit the types and sizes of devices on which the model can be deployed. When selecting algorithms and models, it is important to balance performance requirements, hardware resources, and deployment environments to achieve optimal results in real-world applications. To assist in the study of practical applications, we introduce two parameters as effective references.

To completely analyze the OIDS-45 dataset, we evaluate many of the most powerful and common target identification frameworks now available on the OIDS-45 dataset, comparing their effectiveness to the IP102 dataset tested under identical conditions.

3.1. Experimental Design

We evaluated 12 more advanced or typical target detection methods on the OIDS-45 dataset, and based on the features of the OIDS-45 dataset, we also made targeted improvements to the corresponding network, trying to be more relevant to the actual application situation. We employ both single-stage target detection models, such as YOLOv8 [53], YOLOv7 [54], EffitientDet [55], and RetinaNet [56], and two-stage target detection models, such as Faster R-CNN [57], Mask R-CNN [58], and Cascade R-CNN [59]. detection models; meanwhile, as the target detection performance based on Anchor [60] reaches its limit, target detection algorithms based on Anchor-free [61] have become the current research hotspot, and we therefore use Anchor-free target detection algorithms, such as CenterNet [62], FCOS [63], etc., to evaluate the OIDS-45 dataset.In practice, because the insect dataset has the characteristic of detecting very small objects, and with the wide and hot application of Vision-Transformer [64] in the target detection task, in order to study the detection effect of such small targets as insects in more depth, we also adopted the Focus-DETR [65] and ViT-FRCNN [66] target detection networks to evaluate the dataset, and we also tried to replace the backbone of the YOLOv7 network with Vision-Transformer.Among these, Focus-DETR is an enhancement to the DETR [67] model that focuses on tackling the long-tail problem in target detection. The DETR model is an end-to-end target detection model that implements the target detection job using the Transformer [68] architecture, eliminating difficult phases such as anchor frames, candidate frames, and non-extremely large suppression from standard target detection approaches.In addition, we introduced a semi-supervised goal learning framework called Effecient-Teacher [69] to verify the training effect when the OIDS-45 dataset is incompletely labeled and to test whether the dataset still has a higher accuracy with as little labeling as possible, in which we used YOLOv7 as the detector [70]. Finally, we employed the IP102 dataset in the same scenario for comparison studies, seeking to emphasize the peculiarities of our dataset by comparing it to more common datasets available today.

During the experiments, we trained each network structure for 300 rounds using the Adam [71] optimizer. The batch size was set to 16, and the learning rate [72] was kept at 0.001. The model received images of pixel size 540*540 uniformly during the training task. The target detection experiments were conducted using PyTorch and executed on an NVIDIA 4080 GPU with 8 GB of memory and an Intel server. This ensured the stability and reliability of the experiments while making the best use of computational resources. The specific training parameter settings are shown in Table 4.

Table 4

Specific Model Configuration Parameters
Batch_Size[73]	Learning-rate	pixels	Channel	optimizer
16	0.001	540*540	3	Adam

3.2. Experimental Results and Discussion

As shown in Table 5, through experimentation with the aforementioned network structures, we have observed that our dataset performs better on single-stage target detection models, including YOLOv8, YOLOv7, EfficientDet, and RetinaNet, as well as on anchor-free target detection models such as CentreNet and FCOS. Additionally, our experiments have reaffirmed the powerful role of the vision transformer in the target detection task. Among the compared models, the ViT-FRCNN target detection network, which uses Vision-Transformer as the backbone network, achieves the highest detection accuracy. Our dataset shows better experimental results than the IP102 dataset under the same conditions. This can be attributed to the following reasons:

We have more samples in our dataset for each class of insect.
Our dataset was subjected to the data preprocessing process, such as data enhancement, prior to the experiments.
We simulated the state of the images closer to the actual situation through the data preprocessing.

However, our experimental results do not match the detection performance of current state-of-the-art target detection networks. As can be seen from the confusion matrix in Fig. 8, insects with a multi-stage growth form and insects with similarities in appearance to other categories have lower detection accuracy. In our view, the main reasons for this situation are:

Insects with more than one growth stage are not well detected; As shown in Fig. 6, the coccolithophore moth has a low detection effect, and its prediction accuracy can only reach 72.3% because it has two forms: adults and larvae, and the difference between the two forms is large.
There are several types of insects that are morphologically similar and have poor detection performance; e.g., heterochromatic ladybirds, seven-star ladybirds, and tortoise ladybirds belong to the same family of ladybirds but belong to different classes, with the main differences being patterns and colors.
Insect targets are typically small, and the SSD [74] architecture can be used to detect targets of varying sizes using multi-scale feature maps. However, when dealing with small targets, SSD detection results tend to blur or miss detections. This is because feature maps for small targets are small, resulting in information loss, while features relating to the target object may be veiled or only exist in a few feature channels.

During the 300 rounds of training for the target detection task, the YOLOv8 network showed a sharp decrease in both training and validation loss curves at round 270, indicating significant progress and successful learning of key data features. As depicted in Fig. 7, around round 300, the flattening of the loss curves may indicate that the model is approaching convergence, meaning that the model parameters are no longer changing significantly. We determined that the model had not yet overfit and that further training would likely result in only a small performance gain. We stopped training after 300 rounds of the task to avoid overfitting. We then evaluated and tested the model to confirm its generalizability.

Table 5

Detection performance of different object detection algorithms
Dataset	Method	Backbone	Pre	Recall	mAP	mAP.50	mAP.75
OIDS-45	YOLOv8	DarkNet-53[75]	91.15%	91.12%	72.74%	91.12%.	59.62%
	YOLOv7	Swim-Transformer	92.28%	93.84%	80.12%	94.36%.	66.23%
	EffitientDet	EffitientNet	84.32%	89.27%	69.33%	87.12%.	52.14%
	RetinaNet	ResNet50[76]	90.44%	91.35%	65.31%	90.03%.	55.24%
	Faster R-CNN	ResNet101[77]	68.38%	88.49%	51.32%	80.51%.	39.36%
	Mask R-CNN	ResNet101	71.29%	90.08%	53.26%	82.47%.	43.34%
	Cascade R-CNN	ResNet101	78.32%	92.33%	58.21%	81.65%.	46.28%
	CenterNet	DLA-Net[78]	93.53%	92.12%	81.36%	96.87%.	71.39%
	FCOS	ResNeXt[79]	88.15%	92.31%	76.39%	88.62%.	54.37%
	Focus-DETR	ResNet50	85.43%	87.70%	60.45%	85.32%.	51.29%
	ViT-FRCNN	Vision-Transformer	92.97%	94.68%	81.34%	93.25%.	62.38%
	Effetient-Teacher	YOLOv7 + MobileOne	88.31%	90.02%	66.72%	86.74%.	57.23%
IP-102	YOLOv8	DarkNet-53	82.36%	90.41%	62.39%	84.54%.	51.32%
	YOLOv7	Swim-Transformer	84.27%	93.21%	67.71%	88.31%.	54.13%
	EffitientDet	EffitientNet	74.61%	85.42%	45.62%	64.29%.	32.11%
	RetinaNet	ResNet50	79.68%	88.31%	54.29%	61.77%.	31.26%
	Faster R-CNN	ResNet101	61.31%	84.36%	43.21%	53.01%	21.33%
	Mask R-CNN	ResNet101	63.45%	90.20%	45.12%	67.25%.	25.79%
	Cascade R-CNN	ResNet101	71.15%	93.28%	50.13%	71.07%.	35.31%
	CenterNet	DLA-Net	78.88%	91.29%	61.04%	82.36%.	52.30%
	FCOS	ResNeXt	77.52%	87.41%	59.39%	81.94%.	49.37%
	Focus-DETR	ResNet50	81.37%	91.39%	63.94%	74.77%.	50.32%
	ViT-FRCNN	Vision-Transformer	83.21%	94.65%	69.43%	87.41%.	55.57%
	Effetient-Teacher	YOLOv7 + MobileOne	75.09%	85.92%	59.74%	71.31%.	51.32%

3.3. Challenges and Limitations

Finally, we will discuss the challenges and limitations of the OIDS-45 dataset. Since insects have different morphologies in different life cycles [80], this makes it difficult for target detection and classification models to normalize them into the same category, as it is difficult to extract discriminative features.

However, the imbalance in data distribution in the dataset should not be overlooked. The OIDS-45 dataset reveals an uneven distribution at several levels, as well as a long-tail impact. The long-tail effect in a dataset refers to the presence of a distribution in which a small number of categories or samples account for the majority of the data volume, while the majority of the categories or samples account for a minor portion of the data volume. Figure 9 indicates that in the OIDS-45 dataset, the insect category Chrysoperla sinica has the most photos (1849), whereas the insect Aphanostigma piri only has 149, indicating a clear long-tail effect phenomenon. In the target detection task, the long-tail effect may result in a low number of target samples for some categories in the dataset and a high number of target samples for others, resulting in a data imbalance problem that affects the model's ability to detect targets in low-sample categories. In datasets with long-tailed distributions, for categories with few samples, the model may be unable to properly learn the properties of these categories, resulting in misdetections or omissions in target detection. To summarize, the long-tail effect on the target detection task can cause issues such as data imbalance and model training difficulties, and measures such as data augmentation to increase the number of samples for rare categories should be targeted to address these issues, thereby improving the model's performance and generalization ability.

In this work, we collected a large orchard insect dataset called OIDS-45 for orchard insect identification, including more than 58595 labeled images of 45 species. Compared with previous datasets, OIDS-45 conforms to several characteristics of insect distributions in real environments (e.g., diversity and class imbalance). Also, we evaluated some state-of-the-art identification methods on our dataset. The results show that current deep learning target detection algorithms are not yet able to handle the problem of accurate identification and detection of orchard insects well due to the various characteristics of the insects themselves. We hope that this work will help advance future research on some fundamental problems and common target detection tasks. At the same time, we believe that the large-scale insect dataset of this orchard provides a rich information resource for the target detection task, and in the future, we can consider introducing multimodal data (e.g., sound or infrared images), using reinforcement learning techniques to improve the annotation efficiency, and exploring federated learning for model training across data sources. In addition, combining biological knowledge with deep learning can improve the understanding of entomological features in target detection methods, promote cross-disciplinary research, and advance the development of intelligent monitoring technology in agriculture.

curation, Hongkun Chen, Junyang Chen and Yingjie Xie; Formal analysis, Junyang Chen, Hongkun Chen, Hangfei He, Li Wan and Jingjie Guo; Investigation, Hongkun Chen, Yingjie Xie and Boyi Zhang; Methodology, Hongkun Chen, Yingjie Xie, Li Wan and Xiaoyan Chen; Project administration, Xiaoyan Chen and Hongkun Chen; Resources, Xiaoyan Chen; Software, Hongkun Chen and Junyang Chen; Supervision, Xiaoyan Chen; Validation, Junyang Chen, Qiurui Liu and Jun Li; Visualization, Hongkun Chen; Writing – original draft, Hongkun Chen,Junyang Chen, Yingjie Xie and Li Wan; Writing – review & editing, Hongkun Chen, Junyang Chen and Xiaoyan Chen.

Funding: The project is funded by the National College Student Innovation Training Project. Research Fund of Sichuan Agricultural University, project number: 202310626024

Data Availability Statement: The datasets generated and analysed during the current study are available in the Github repository, https://github.com/XieYingjie66/OIDS-45.

Acknowledgments: We would like to thank the School of Information Engineering at Sichuan Agricultural University and Ms. Wan Li for their strong support of the research project.

Conflicts of Interest: The authors declare no conflict of interest.

Helvacı, Murat. "Insect Pest Management in Fruit Production." Fruit Industry. IntechOpen, 2022.
Dent, David, and Richard H. Binks. Insect pest management. Cabi, 2020.
Zhang, Qingsong, et al. "Outlook of China's agriculture transforming from smallholderoperation to sustainable production." Global Food Security 26 (2020): 100444.
Cai, yang, et al. "Pesticide overuse in apple production and its socioeconomic determinants: Evidence from Shaanxi and Shandong provinces, China." Journal of CleanerProduction 315 (2021): 128179.
Wu, et al. "Ip102: A large-scale benchmark dataset for insect pest recognition." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2019.
Baker, Brian P., Thomas A. Green, and Ali J. Loker. "Biological control and integratedpest management in organic and conventional systems." Biological Control 140 (2020):104095.
Poveda, Jorge. "Insect frass in the development of sustainable agriculture. A review." Agronomy for Sustainable Development 41.1 (2021): 5.
Tong, Yi-Jie, et al. "The relationship between genus/species richness and morphological diversity among subfamilies of jewel beetles." Insects 12.1 (2021): 24.
Altaheri, Hamdi, et al. "Deep learning techniques for classification of electroencephalogram (EEG) motor imagery (MI) signals: A review." Neural Computing and Applications 35.20 (2023): 14681–14722.
Wu, Yue, et al. "Rethinking classification and localization for object detection." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2020.
Pal, Sankar K., et al. "Deep learning in multi-object detection and tracking: state of the art." Applied Intelligence 51 (2021): 6400–6429.
Li, Wei, et al. "Summary of target detection algorithms." Journal of Physics: Conference Series. Vol. 1757. No. 1. IOP Publishing, 2021.
Chandra, Mayank Arya, and S. S. Bedi. "Survey on SVM and their application in image classification." International Journal of Information Technology 13 (2021): 1–11.
Lei, and Ye Yao. "Study on an adaptive thermal comfort model with K-nearest-neighbors (KNN) algorithm." Building and Environment 202 (2021): 108026.
Shahraki, Amin, Mahmoud Abbasi, and Øystein Haugen. "Boosting algorithms for network intrusion detection: A comparative evaluation of Real AdaBoost, Gentle AdaBoost and Modest AdaBoost." Engineering Applications of Artificial Intelligence 94 (2020):103770.
Lu, Yang, et al. "Identification of rice diseases using deep convolutional neural networks." Neurocomputing 267 (2017): 378–384.
Hershey, Shawn, et al. "CNN architectures for large-scale audio classification." 2017 ieee international conference on acoustics, speech and signal processing (icassp). IEEE,2017.
Wang, Jiangning, et al. "A new automatic identification system of insect images at the order level." Knowledge-Based Systems 33 (2012): 102–110.
Wei, Guangfen, et al. "Development of a LeNet-5 gas identification CNN structure for electronic noses." Sensors 19.1 (2019): 217.
Iandola, Forrest N., et al. "SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and < 0.5 MB model size." arv preprint arv:1602.07360 (2016).
Yang, Guoguo, Yidan Bao, and Ziyi Liu. "Localization and recognition of pests in tea plantation based on image saliency analysis and convolutional neural network." Transactions of the Chinese Society of Agricultural Engineering 33.6 (2017): 156–162.
Thenmozhi, K., and U. Srinivasulu Reddy. "Crop pest classification based on deep convolutional neural network and transfer learning." Computers and Electronics in Agriculture 164 (2019): 104906.
Xie, Chengjun, et al. "Multi-level learning features for automatic classification of fieldcrop pests." Computers and Electronics in Agriculture 152 (2018): 233–241.
Zhong, Yuanhong, et al. "A cascade reconstruction model with generalization ability evaluation for anomaly detection in videos." Pattern Recognition 122 (2022): 108336.
Rice, Leslie, Eric Wong, and Zico Kolter. "Overfitting in adversarially robust deep learning." International Conference on Machine Learning. PMLR, 2020.
Yang, Yao-Yuan, et al. "A closer look at accuracy vs. robustness." Advances in neuralinformation processing systems 33 (2020): 8588–8601.
Andrekha, Muhammad Zikri, and Yasdinul Huda. "Deteksi Warna Manggis Menggunakan Pengolahan Citra dengan Opencv Python." Voteteknika (Vocational Teknik Elektronika dan Informatika) 9.4 (2021): 27–33.
Wu, Geming, Shuqian Luo, and Zhi Yang. "Optimal weighted bilateral filter with dual-range kernel for Gaussian noise removal." IET Image Processing 14.9 (2020): 1840–1850.
DeVries, Terrance, and Graham W. Taylor. "Improved regularization of convolutional neural networks with cutout." arv preprint arv:1708.04552 (2017).
Dadboud, Fardad, et al. "Single-stage uav detection and classification with yolov5: Mosaic data augmentation and panet." 2021 17th IEEE International Conference on Advanced Video and Signal Based Surveillance (AVSS). IEEE, 2021.
Choudhury, Nurul Amin, and Badal Soni. "An Adaptive Batch Size based-CNN-LSTM Framework for Human Activity Recognition in Uncontrolled Environment." IEEE Transactions on Industrial Informatics (2023).
Ioffe, Sergey, and Christian Szegedy. "Batch normalization: Accelerating deep network training by reducing internal covariate shift." International conference on machine learning. pmlr, 2015.
Gai, Rongli, Na Chen, and Hai Yuan. "A detection algorithm for cherry fruits based on the improved YOLO-v4 model." Neural Computing and Applications 35.19 (2023): 13895–13906.
Zhang, Pengyi, et al. "CoSinGAN: learning COVID-19 infection segmentation from a single radiological image." Diagnostics 10.11 (2020): 901.
Shaham, Tamar Rott, Tali Dekel, and Tomer Michaeli. "Singan: Learning a generative model from a single natural image." Proceedings of the IEEE/CVF international conference on computer vision. 2019.
Cheng, Yu, et al. "Sequential attention GAN for interactive image editing." Proceedings of the 28th ACM international conference on multimedia. 2020.
Cubuk, Ekin D., et al. "Randaugment: Practical automated data augmentation with a reduced search space." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. 2020.
Yakovlev, A., and O. Lisovychenko. "An approach for image annotation automatizationfor artificial intelligence models learning." Адаптивні системи автоматичного управління 1.36 (2020): 32–40.
RK Samanta and Indrajit Ghosh. Tea insect pests classification based on artificial neural networks. International Journal of Computer Engineering Science, 2(6):336, 2012.
XJiangning Wang, Congtian Lin, Liqiang Ji, and Aiping Liang. A new automatic identification system of insect images at the order level. Knowledge-Based Systems, 33:102–110, 2012.
Kanesh Venugoban and Amirthalingam Ramanan. Image classification of paddy field insect pests using gradient-based features. International Journal of Machine Learning and Computing, 4(1):1–5, 2014.
Chengjun Xie, Jie Zhang, Rui Li, Jinyan Li, Peilin Hong, Junfeng Xia, and Peng Chen. Automatic classification for field crop insects via multiple-task sparse representation and multiple-kernel learning. Computers and Electronics in Agriculture, 119:123–132, 2015.
Ziyi Liu, Junfeng Gao, Guoguo Yang, Huan Zhang, and Yong He. Localization and classification of paddy field pests using a saliency map and deep convolutional neural network. Scientific Reports, 6:20410, 2016.
Denan, et al. "Insect detection and classification based on an improved convolutional neural network." Sensors 18.12 (2018): 4169.
Ji, Lei, et al. "How much can AI techniques improve surface air temperature forecast?—A report from AI Challenger 2018 Global Weather Forecast Contest." (2019): 989–992.
Wu,, et al. "Ip102: A large-scale benchmark dataset for insect pest recognition." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2019.
Limiao Deng, Yanjiang Wang, Zhongzhi Han, and Renshi Yu. Research on insect pest image detection and recognition based on bio-inspired methods. Biosystems Engineering,169:139–148, 2018.
Mazen, Fatma MA. "Arthropod Taxonomy Orders Object Detection in ArTaxOr dataset using YOLOX." Journal of Engineering and Applied Science 70.1 (2023): 113.
Salehin, Imrus, et al. "Deep Learning Models for Real-Time Identification of Agro Plant Insects with Image Augmentation Effect Analysis on a Small Custom Dataset." Available at SSRN 4619643.
Khan, Shahbaz, et al. "A novel framework for multiple ground target detection, recognition and inspection in precision agriculture applications using a UAV." Unmanned Systems 10.01 (2022): 45–56.
Peng, gchao, et al. "CourtNet: Dynamically balance the precision and recall rates in infrared small target detection." Expert Systems with Applications 233 (2023): 120996.
Qi, Shengxiang, et al. "Robust method for infrared small-target detection based on Boolean map visual theory." Applied optics 53.18 (2014): 3929–3940.
Lou, Haitong, et al. "DC-YOLOv8: Small-Size Object Detection Algorithm Based on Camera Sensor." Electronics 12.10 (2023): 2323.
Wang, Chien-Yao, Alexey Bochkovskiy, and Hong-Yuan Mark Liao. "YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023.
Koonce, Brett, and Brett Koonce. "EfficientNet." Convolutional Neural Networks with Swift for Tensorflow: Image Recognition and Dataset Categorization (2021): 109–123.
Wang, Yuanyuan, et al. "Automatic ship detection based on RetinaNet using multi-resolution Gaofen-3 imagery." Remote Sensing 11.5 (2019): 531.
Girshick, Ross. "Fast r-cnn." Proceedings of the IEEE international conference on computer vision. 2015.
He, Kaiming, et al. "Mask r-cnn." Proceedings of the IEEE international conference oncomputer vision. 2017.
Cai, Zhaowei, and Nuno Vasconcelos. "Cascade r-cnn: Delving into high quality objectdetection." Proceedings of the IEEE conference on computer vision and pattern recognition. 2018.
Kong, Tao, et al. "Foveabox: Beyound anchor-based object detection." IEEE Transactions on Image Processing 29 (2020): 7389–7398.
Tian, Zhi, et al. "FCOS: A simple and strong anchor-free object detector." IEEE Transactions on Pattern Analysis and Machine Intelligence 44.4 (2020): 1922–1933.
Duan, Kaiwen, et al. "Centernet: Keypoint triplets for object detection." Proceedings of the IEEE/CVF international conference on computer vision. 2019.
Kannadaguli, Prashanth. "FCOS Based Human Detection System Using Thermal Imaging for UAV Based Surveillance Applications." 2020 IEEE Bombay Section Signature Conference (IBSSC) IEEE, 2020.
Han, Kai, et al. "A survey on vision transformer." IEEE transactions on pattern analysis and machine intelligence 45.1 (2022): 87–110.
Zheng, Dehua, et al. "Less is more: Focus attention for efficient detr." Proceedings ofthe IEEE/CVF International Conference on Computer Vision. 2023.
Beal, Josh, et al. "Toward transformer-based object detection." arv preprint arv:2012.09958 (2020).
Dai, yang, et al. "Dynamic detr: End-to-end object detection with dynamic attention." Proceedings of the IEEE/CVF International Conference on Computer Vision. 2021.
Han, Kai, et al. "Transformer in transformer." Advances in Neural Information Processing Systems 34 (2021): 15908–15919.
Xu, Bowen, et al. "Efficient Teacher: Semi-Supervised Object Detection for YOLOv5." arv preprint arv:2302.07577 (2023).
Li, Zeming, et al. "Detnet: A backbone network for object detection." arv preprint arv:1804.06215 (2018).
Zhang, Zijun. "Improved adam optimizer for deep neural networks." 2018 IEEE/ACM 26th international symposium on quality of service (IWQoS). Ieee, 2018.
Zeiler, Matthew D. "Adadelta: an adaptive learning rate method." arv preprint arv:1212.5701 (2012).
Hu, Zhongzhe, et al. "Fast and accurate variable batch size convolution neural network training on large scale distributed systems." Concurrency and computation: practice and experience (2022)..
Sehwag, Vikash, Mung Chiang, and Prateek Mittal. "Ssd: A unified framework for self-supervised outlier detection." arv preprint arv:2103.12051 (2021).
Yang, Lina, Gang Chen, and Wenyan Ci. "Multiclass objects detection algorithm usingDarkNet-53 and DenseNet for intelligent vehicles." EURASIP Journal on Advances in Signal Processing 2023.1 (2023): 85.
Theckedath, Dhananjay, and R. R. Sedamkar. "Detecting affect states using VGG16, ResNet50 and SE-ResNet50 networks." SN Computer Science 1 (2020): 1–7.
Zhang, Qi. "A novel ResNet101 model based on dense dilated convolution for imageclassification." SN Applied Sciences 4 (2022): 1–13.
Su, Yanfei, et al. "DLA-Net: Learning dual local attention features for semantic segmentation of large-scale building facade point clouds." Pattern Recognition 123 (2022): 108372.
Zhou, Tianyan, Yong Zhao, and Jian Wu. "Resnext and res2net structures for speakerverification." 2021 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2021.
Nylin, Sören. "Life history perspectives on pest insects: What’s the use?." Austral Ecology 26.5 (2001): 507–517.

No competing interests reported.

CategoryDistribution.png

Download PDF

Version 1

posted

You are reading this latest preprint version

OIDS-45: A large-scale benchmark insect dataset for orchard pest monitoring

Status:

Version 1

Abstract

Figures

1. Introduction

2. Materials and Methods

2.1. Related work

2.1.1 Insect identification technology

2.1.2 Relevant datasets

2.2. Creation of datasets

2.2.1. Preliminary research work

2.2.2. Image collection

2.2.3. Data filtering and selection

2.2.4. Data preprocessing

2.2.5. Professional data annotation

2.3. Dataset partitioning

2.4. Dataset structure

2.5. Comparison with other datasets

2.6. Evaluation Metrics

3. Results and Discussion

3.1. Experimental Design

3.2. Experimental Results and Discussion

3.3. Challenges and Limitations

4. Conclusions

Declarations

References

Additional Declarations

Supplementary Files

Status:

Version 1