Intelligent solid waste classification system using combination of image processing and machine learning models

doi:10.21203/rs.3.rs-2573812/v1

Download PDF

Research Article

Intelligent solid waste classification system using combination of image processing and machine learning models

https://doi.org/10.21203/rs.3.rs-2573812/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Solid waste is a major issue in all countries around the world. Solid waste classification and segregation prior to reuse, recycling or recovery is an important step toward sustainable waste management. Traditional manual sorting of solid waste is a labour intensive process that may pose health risks to the workers. Currently, automated classification of solid waste using machine learning techniques are widely applied. This study is aiming to develop an automated waste classification model by testing traditional and deep machine learning models. To achieve that, both open (Trashnet) and generated datasets were used in the model training and testing. The study results showed relatively low prediction capability of the traditional machine learning models like Random Forest (RF) and Support Vector Machine (SVM) as compared to the deep machine learning Convolutional Neural Network (CNN). The testing of the three models on a combined data set of Trashnet with local garbage data set resulted in accuracy of 62.5% for SVM, 72.0% for RF and 92.7% for CNN.

JONET deep learning model has been developed using a combination of pre-trained base model (DenseNet 201) with a new architicture that contains a fully connected layer in the classification stage with 1024 neurons. The model is capable to identify six classes of solid waste items with various accuracies. When tested on the Trashnet, the accuracy was 96.06%, while testing on the local garbage dataset gave an accuracy of 94.40%. JONET has been tested also on multi object images which gave an acceptable prediction accuracy.

Artificial Intelligence

Image processing

Solid waste

CNN

Deep learning

Solid waste management is a challenge that faces many cities around the world. Integrated solid waste management strategies, like reduce, reuse, recycling, composting, and waste-to-energy, play a significant role in reducing the adverse impacts of solid waste landfilling on the environment and human health (Memon, 2010, Abu Qdais et al., 2019). Availability of reliable data on solid waste composition is a vital step before deciding which resource recovery option to follow (Abu Qdais, 2007, Dahlen and Lagerkvist 2008, Edjabou et al., 2015). Traditionally, manual sorting is the most commonly used method in determining the composition of solid waste. However, manual sorting is a labor intensive that requires larger space (Vrancken et al., 2017) and long duration (Tao et al., 2023) and may pose health risks to the personnel involved in the sorting process (Wilts et al., 2021). To minimize the adverse impacts of the manual sorting, a paradigm shift towards more intelligent sorting and classification models started taking place recently (Adeleke et al., 2021).

Image Processing (IP) and Artificial Intelligence (AI) are promising techniques in the field of urban planning, which can contribute to achieve the sustainable city agenda (Vrancken et al., 2017 Casali et al., 2022). This is particularly true in the field of solid waste management, where the adoption of IP and AI has resulted in significant cost savings for the various stakeholders involved in the SWM chain (Varncken et al., 2017, Wilts et al., 2021), including collection, sorting and resource recovery which has led to environmental improvements on multiple levels (Helm et al., 2020, Lin et al., 2022), and could pave the way to enable the adoption of circular economy options (Wilts and Berg, 2018).

Recently, several researchers used IP and AI in solid waste management. Abu Qdais and Shatnawi (2019) merged remote sensing and machine learning techniques in estimating the landfill surface temperature. Yang and Thung (2016) employed a support vector machine (SVM) with scale-invariant feature transform (SIFT) and a convolutional neural network (CNN) to categorize photos of a single waste object into six distinct categories. Another study by Mao et al. (2021) used the TrashNet Dataset, which comprises 2527 images with six different waste categories, to evaluate the CNNs' performance.

In an attempt to identify and classify different items in the solid waste stream based on TrashNet dataset, Meng and Chu (2020) used different models including support vector machines (SVM) with HOG features, which gave an accuracy of 47.25% and 23.51% with and without augmentation, respectively, while the convolutional neural network (CNN) achieved a test accuracy of 93% using a 90/10 training/testing augmented data split. Bircanoglu et al. (2018) tested the well-known deep convolutional neural network architectures using the TrashNet dataset. The researchers reported that Inception-Resnet, and Inception-v4 outperformed all others with 90% test accuracy, while for transfer learning and fine-tuning of weight parameters using ImageNet, DenseNet121 gave the best result with 95% test accuracy on TrashNet Dataset.

By combining a pre-trained CNN with a network header to classify TrashNet images taken by mobile Shi et al. (2020) achieved an image classification accuracy of 93.68%. Another study by Xu et al. (2020) presented a waste classification and identification method based on transfer learning and a lightweight neural network. By migrating the lightweight neural network MobileNetV2 and rebuilding it, the reconstructed network is used for feature extraction. The model was trained and verified using 2527 images of garbage-labeled data in the TrashNet dataset, resulting in a classification accuracy of 98.4%. Guo et al. (2021) reported that the most widely used machine learning models in municipal solid waste studies is the artificial neural networks (ANN) followed by support vector machine (SVM) models.

The main objective of the present study is to develop an automated recognition system that is capable to recognise various items in the solid waste stream using a combination of image processing and artificial intelligence techniques based on both single and multiple objects images.

2.1 Solid waste input data sets

In this research collected images of municipal solid waste were subjected to analysis. A total of 5151 RGB-coded images were obtained from two sources. First source is the TrashNet data set (Yang and Thung, 2016) where all the available 2495 images were used in the study. Deep learning models require larger number of data input (Shi et al. 2020). Hence, to develop a larger dataset the TrashNet data set was supplemented by additional set of images (Rajak et al., 2020) that was obtained by taking 2656 shots of the local garbage in Jordan using a digital camera. Six items of solid waste were classified, namely paper, cardboard, glass, metals, plastics and miscellaneous trash. Table 1 presents the characteristics of the datasets utilized in developing the models. As it can be observed from the table, there are four groups of datasets used in the development of the machine learning models. Models that used solely the TrashNet or local garbage images and models that used combination of both sets (TrashNet and local images). In all these cases the models were developed based on a single solid waste item in each image. To reflect the real case where the solid waste items are usually mixed in one stream, another dataset was created by shooting images of multiple items of the local solid waste, where two or more items are available in each image.

Table 1

Datasets used in the model development
Data set name	Number of images	Solid waste items and number of images of each item
TrashNet*	2495	Cardboard	438
		Glass	500
		Metal	405
		Paper	582
		Plastic	496
		Trash	74
Local garbage dataset	2656	Cardboard	635
		Glass	481
		Metal	447
		Paper	349
		Plastic	610
		Miscellaneous	134
TrashNet dataset merged with the local garbage dataset	5151	Cardboard	1073
		Glass	981
		Metal	852
		Paper	931
		Plastic	1106
		Miscellaneous	208
Multiple objects local garbage dataset	200	Two or more classes were found in each image.

2.2 Image pre-processing

Variation in brightness or availability of unwanted signals in each image that is referred to as a noise. Such a noise may corrupt the images and reduce their accuracy (Owotogbe et al, 2019). Therefore, before being introduced as input to the machine learning, the raw images in both data sets were subjected to preprocessing to render them suitable as input to the developed models. To achieve that, all images were resized, and rescaled using a scale of 1:255.

2.3 Image Augmentation and Segmentation

The larger the datasets used in developing the models, the better the results of the machine learning models. To increase the size and quality of the datasets used, image data augmentation is adopted (Shorten and Khoshgoftaar, 2019). In this study the dataset augmentation was performed in both the pre-training as well as during the training phases. In the pre-training phase augmentation of the dataset was accomplished using imageuag library (Schnieder et al, 2020) and simple geometric transforms from Keras (Nanez et al, 2021), while during the training phase the augmentation took place by using Keras image data generator tool. For the images with multiple objects, a segmentation process was adopted using a Canny edge technique which performs better than other techniques (Maini and Aggarwal, 2009). After the segmentation the images were introduced to the machine learning model as single object images.

2.4 Machine learning models

To achieve the objectives of this research, several machine learning models were developed and tested, including traditional as well as deep learning algorithms. This includes Support Vector Machine(SVM), Random Forest (RF), and Deep CNN. These models were compared based on several metrics including, F1 score, Precision, Recall and Accuracy.

Since it has several basic and effective tools for machine learning, Python programing language is among the most frequently used ones solve diverse problems (Srinath,2017). Therefore, in his study, Python has been utilized in the model’s development.

Figure 1 illustrate a flowchart of the methodology followed in conducting the study. To develop a local dataset of solid waste images, the solid waste was obtained from street containers, where images of each item been taken using a digital camera. Models have been trained and tested on the collected datasets.

For developing the models, the datasets were divided into two groups, where 80% of the data used for model training and 20% (Niu et al., 2022) for testing the original data without augmentation. On the other hand, when augmentation was adopted, the dataset was splitted into 80% for model training, 10% for validation and 10% for testing (Marin et al., 2021).

The study compares the efficiency and accuracy of CNN models with two other supervised learning algorithms, namely Support Vector Machines (SVM) and Random Forest (RF). To determine the optimum parameters for the traditional models development, GridSearch algorithm was used from the Sk-Learn interfaces for hyperparameter tuning. For Random Forest (RF), it assumed that the most critical hyperparameters were the number of estimators (n) and the criterion (gini or entropy). A value of 200 was chosen for the number of trees in the forest (n), and the gini optimizer was chosen as the tree criterion, as the average training time and cross-validation value are both acceptable. The results showed that the model's accuracy was 72% with a training time of 0.142 minutes and a standard deviation of 0.08. On the otherhand, for the SVM model the accuracy was 63% with a training time of 2.43 minutes. Such relatively low accuracies of the traditional models may be attributed to the fact that these models are not suitable for classification of solid waste items from RGB images. This is in agreement with the findings of Shi et al., 2021 who reported that traditional machine learning algorithems cannot fit the data and balance the sample well. This is also confirmed by Sothe et al., (2020) who indicated that the performance of CNN is more accurate than SVM and RF. Therefore, a CNN model has been developed in this study. The model included both local images dataset and TrashNet datasets for both single and multiple images (Mao et al., 2022). Since the images of the solid waste generated in Jordan used in the model development, the model has been called JONET model.

3.1 JONET Model

The use of CNN models in image processing and recognition has been applied in many areas (Bobulski and Kubanek, 2019) including solid waste classification (Shi et al, 2021, Xia et al., 2021, Malik et al., 2022) and inteligent waste identification and recycling (Wu et al., 2023). The concept of transfer learning (Wu et al., 2023) was used in this research to transfer weights from pre-trained models and then reset them to be compatible with the local garbage and TrashNet datasets. The fine-tuning concept was also used by removing the old classifier for pre-trained models and replacing it with a new one that suited the dataset (Kaya et al., 2019).

To decide on the proper model architecture, variety of different fully connected layers with different numbers of neurons were tested and analyzed. Figure 2a. illustrates the selected architicture of the JONET model. As shown, the model consists of two blocks, the first block is dealing with the feature extraction, while the second one utilized for image classification. The feature extraction process was accomplished by unfreezing all convolutional layers of the pre-trained base model. After that, the extracted features were introduced into a fully connected layer that contains 1024 neurons for classification of the solid waste items. Sixteen base models were tested and evaluated as shown in Fig. 2.b. As it can be seen, DenseNet 201 gave the highest accuracy of 92.7% among all the tested pre-trained models with only 67 epochs, which makes it suitable to be selected in the classification stage of model development. Various researchers adopted different pre-trained base models. For example, Kryzhanovsky et al. (2020) introduced a method for selecting a pre-trained CNN for fine tuning on a fresh image sample. The method is based on estimating the separability of pre-trained CNN features for the examined image classification issue and selecting the CNN with the highest one. The selection of the suitable base model is relevant to several factors that may include the CNN model architecture, dataset size and characteristics, where CNN complexity is not always reflecting a higher accuracy of the model (Choromanski et al., 2020).

The performance of the developed deep learning algorithm for classifying the images was evaluated using the performance indices (Chowdhury, et al, 2020) namely, recall, precision, and F1 score as given by equations 1 to 3. Precision refers to how the ratio of waste item was correctly predicted, while recall is the accuracy ratio of waste item predicted and how they matched to categories. These indices must be as high as possible.

$$\text{A}\text{c}\text{c}\text{u}\text{r}\text{a}\text{c}\text{y}=\frac{\text{T}\text{P}+\text{T}\text{N}}{\left(\text{T}\text{P}+\text{F}\text{N}\right)+(\text{F}\text{P}+\text{T}\text{N})}$$

$$\text{P}\text{r}\text{e}\text{c}\text{i}\text{s}\text{i}\text{o}\text{n}=\frac{\text{T}\text{P}}{\text{T}\text{P}+\text{F}\text{P}}$$

$$\text{F}1 \text{S}\text{c}\text{o}\text{r}\text{e}=\frac{2\text{*}\text{T}\text{P}}{2\text{*}\text{T}\text{P}+\text{F}\text{N}+\text{F}\text{P}}$$

Where, (TP) is true positive, (TN) is true negative, (FP) is false positive and (FN) is false negative.

The confusion matrix and classification report for the JONET Model tested on TrashNet datasets are presented in Table 2. It can be observed that glass has the highest prediction ratio with F1 score of 0.98, where out of 101 images, 100 images were correctly predicted (99.0%) followed by paper 98.3% then metal 97.6%, while the plastics category has the lowest prediction accuracy of 91.8%. The overall accuracy of the model by Eq. (1) is 96.06% with training time of 28 min. Using a four-layer CNN and a total of 400 images, Altikat et al, 2022 reported a lower prediction accuracy for all items, which ranged from 56.7% for plastic to 76.7% for organics. Such a relatively lower accuracy may be attributed to the relatively low number of images used in training and testing the model. This indicates that the number of images used in the training and testing of the CNN model is an important factor in determining the accuracy of prediction. For example, Niu et al., (2022) reported accuracy of 90.62% using 3680 images. Another study by Togacar et al. (2020) CNN with 13966 images of organic class and 11,111 images of recyclables class, achieved an accuracy that reached 99.95%.

Table 2

Confusion Matrix and classification report for JONET tested on TrashNet dataset.
		Predicted Classes
		Cardboard	Glass	Metal	Paper	Metal	Misc.	Total	Precision	F1 Score
True Classes	Cardboard	75	0	0	5	0	1	81	92.6	96
	Glass	0	100	0	0	1	0	101	99.0	98
	Metal	0	2	80	0	0	0	82	97.6	97
	Paper	0	0	1	117	0	1	119	98.3	96
	Plastic	0	2	2	1	89	3	97	91.8	95
	Misc.*	0	0	0	1	0	27	28	96.4	90
	Total	75	104	83	124	90	32	508
	Recall	100.0	96.2	96.4	94.4	98.9	84.4
*Miscellaneous

Using the local garbage dataset alone in JONET also gave relatively high results as shown in Table 3. The results showed that paper class has the highest f1 score of 98% with a precision of 96%. While Plastic has the lowest f1 score of 92% with 87.5% precision. The analysis showed 94.42% of overall accuracy and a training time of 59 minutes. Mao et al. (2022) reported an accuracy of 92.12% by testing on local garbage in Taiwan with training time of 12 hours.

Table 3

Confusion matrix and classification report for JONET tested on local garbage dataset
		Predicted Classes
		Cardboard	Glass	Paper	Plastic	Misc.*	Metal	Total	Precision	F1 Score
True Classes	Cardboard	60	0	0	0	0	0	60	100.0	96
	Glass	2	120	2	0	0	1	125	96.0	98
	Paper	0	1	91	1	0	2	95	95.8	96
	Plastic	1	0	0	28	3	0	32	87.5	92
	Misc.*	1	0	1	0	98	3	103	95.1	92
	Metal	1	0	1	0	10	111	123	90.2	93
	Total	65	121	95	29	111	117	538
	Recall	92.3	99.2	95.8	96.6	88.3	94.9
*Miscellaneous

Table 4 presents the results of the JONET Model classification report and confusion matrix for tests on both datsets (TrashNet and local garbage datasets). It can be noticed that using the combined datasets, the paper class has the greatest prediction ratio, with a f1 score of 0.95, where 175 out of 182 images were likely predicted (96.2%). The metal class has the lowest accuracy of 75% accuracy and f1 score of 0.8 with 30 minutes of training, the model total accuracy is 92.73%. The relatively low accuracy of the model on the combined dataset as compared with the accuracy of testing on individual data sets may be explained by the condition of the solid waste item in the two datasets. The TrashNet data set items are clean ones, while the items in the local dataset have some dirt, as they have collected from the street containers. Similar results were reported by Wu et al. (2023) who indicated that CNN are reliable in identifying the recyclable solid waste items with an accuracy ranged from 70–99.6%. However, when the TrashNet data set extended by additional dataset either from internet or generated locally by the researchers, the prediction accuracy ranged between 79.27% and 96.5%.

Table 4

Confusion matrix and classification report for JONET tested on TrashNet and local garbage datasets.
		Predicted Classes
		Cardboard	Paper	Glass	Plastic	Metal	Misc.	Total	Precision	F1 Score
True Classes	Cardboard	193	8	1	3	2	6	213	90.6	93
	Paper	5	175	0	1	1	0	182	96.2	95
	Glass	0	0	191	9	0	2	202	94.6	94
	Plastic	1	1	9	206	3	5	225	91.6	92
	Metal	1	2	2	3	33	3	44	75.0	80
	Misc.*	0	2	3	2	0	158	165	95.8	93
	Total	200	188	206	224	39	174	1031
	Recall	96.5	93.1	92.7	92.0	84.6	90.8
*Miscellaneous

Table 5. summarizes the findings of the tested models in this study. As it can be seen, the traditional machine learning models, namely SVM and RF are having the lowest accuracies among the tested models with an accuracy of 62.5% and 72% respectively. As for the deep learning models, it can be observed that the highest accuracy was 96.06% for JONET model tested on the TrashNet dataset with a training time of 28 minutes. On the other hand, testing the JONET model on the combined datasets of TrashNet and local images revealed to get the lowest accuracy of 92.73% and a training time of 30 minutes. While using the local garbage images dataset alone resulted in an accuracy of 94.42% with training time of 17.67 minutes.

Table 5

Summary of different machine learning models accuracy tested in the study
Model number	Model	Dataset	Accuracy	Splitting of Data	Training time (min)	Epoch	Input Image Size
1	Support Vector Machine(SVM)	TrashNet + Local Garbage	62.5	Train: Test 80:20	2.43	NA*	32x32
2	Random Forest (RF)	TrashNet + Local Garbage	72.0	Train: Test 80:20	0.142	NA*	32x32
3	JONET model	TrashNet + Local Garbage	92.73	Train: Test 80:20	30	67	256x256
4	JONET model	Local Garbage	94.42	Train: Test 80:20	17.67	59	256x256
5	JONET model	TrashNet	96.06	Train: Test 80:20	28	97	256x256
* not applicable

3.4 Multiple Object Model

In the real life conditions where segregation of solid waste at source is not practiced, the solid waste items are mixed together. Therefore, a models for detecting images with multiple solid waste items are needed in this case, which is rarely addressed in the literature (Mao et al., 2022).

To enhance the capabilities of the CNN in detecting classes of solid waste from multiple object images a multiple object model was developed. Segmentation of the items from the multiobject image was implemented (Mazzeo et al., 2019, Melinto et al., 2020) by splitting the images into single objects. This was achieved by using Canny Edge algorithm. After that the images of each item introduced into the single classification model. Figure 3.a shows the steps of extracting each object from the image by identifying the waste item using a bounding box, while Fig. 3.b is an example of the prediction accuracy of the model using the segmentation technique with JONET model.

In certain cases, the model has made wrong predictions between glass and plastic. This is may be attributed to the transparent nature of both items. Furthermore, it has been observed that the boundary boxes for detecting an object was interfering with each other, where sometimes the object was not clearly identified. Similar problem was reported by Mitra (2020) which was eliminated by decreasing the threshold value during the model testing. The results from the current study indicate the need for further research to develop reliable models for predicting solid waste items from multiple object images.

Nowadays, image processing and artificial intelligent systems are widely applied in the field of solid waste management. Of special importance, are the automated classification and sorting systems of the solid waste that utilize machine learning algorithms. Adopting such systems will minimize the cost and the risks that are associated with the traditional sorting methods. This study compared the traditional (SVM and RF) with a deep (CNN) machine learning models in classifying the solid waste using both TrashNet and local images datasets. Based on the prediction accuracy, CNN remarkably outperformed SVM and RF.

A newly CNN model (JONET) was created jointly based on a combination of pre-trained models with a newly developed architecture. Testing JONET on Trashnet, local dataset and combination of the two datasets resulted in average accuracies of 96.06%, 94.42% and 92.73%, respectively. In this research, it has been demonstrated that JONET model was capable to classify waste items from multi object images using segmentation into single images. Despite the wrong predictions between plastic and glass in some cases, the model performance accuracy is acceptable. The JONET architecture developed in this research can be used to classify solid waste generated in other countries with similar characteristics like that of Jordan.

The reported CNN models in the literature have used different base models, datasets and architecture, which makes the comparison of the findings a hard task. Therefore, there is a pressing need for standardization of modelling process, so as to enable benchmarking. Furthermore, it is recommended to expand the research to develop models based on multiple object images in order to reflect the real situation of the mixed solid waste items when segregation is not adopted.

Abu Qdais HA (2007) Techno-economic assessment of solid waste management in Jordan. Waste Manag 27:1666–1672. 10.1016/j.wasman.2006.08.004
Abu Qdais HA, Shatnawi N (2019) Assessing and predicting landfill surface temperature using remote sensing and an artificial neural network. Int J Remote Sens 40(24):9556–9571
Abu Qdais H, Wuensch C, Dornack C, Nassour A (2019) The role of solid waste composting in mitigating climate change in Jordan. Waste Manag Res 37(8):833–842
Adeleke O, Akinlabi SA, Jen TC, Dunmade I (2021) Application of artificial neural networks for predicting the physical composition of municipal solid waste: An assessment of the impact of seasonal variation. Waste Manage Res 39:1058–1068. 10.1177/0734242X21991642 journals.sagepub.com/home/wmr
Altikat A, Gulbe A, Altikat S (2022) Intelligent solid waste classification using deep convolutional neural networks. Int J Environ Sci Technol 19:1285–1292
Bircanoğlu C, Atay M, Beşer F, Genç Ö, Kızrak MA (2018) July. RecycleNet: Intelligent waste sorting using deep neural networks. In 2018 Innovations in intelligent systems and applications (INISTA) (pp. 1–7). IEEE
Bobulski J, Kubanek M (2019) Waste Classification System Using Image Processing and Convolutional Neural Networks, Proceedings of the 15th International Work-Conference on Artificial Neural Networks, IWANN 2019 Gran Canaria, Spain, June 12–14, 2019, Part II, pp 350–361
Casali Y, Aydin YN, Comes T (2022) Machine learning for spatial analyses in urban areas: a scoping review, Sustainable Cities and Society, 85 (2022) 104050, doi:10.1016/j.scs.2022.104050
Choromanski K, Likhosherstov V, Dohan D, Song X, Gane A, Sarlos T, Weller A (2020) Rethinking attention with performers. arXiv preprint arXiv:2009.14794,. doi.org/10.48550/arXiv.2009.14794
Dahlen L, Lagerkvist A (2008) Review methods for household waste composition studies. Waste Manag 28:1100–1112. 10.1016/j.wasman.2007.08.014
Edjabou ME, Jensen MB, Gotze R, Pivenko K, Peterson C, Scheutz C, Astrup TF (2015) Municipal solid waste composition: Sampling methodology, statistical analyses, and case study evaluation. Waste Manag 36:12–23. doi.org:10.1016/j.wasman.2014.11.009
Chowdhury MEH, Rahman T, Khandakar A, Mazhar R, Abdul Kadir M, Mahbub Z, Khandakar RI, Khan MS, Iqbal A, Al-Emadi N, Reaz MI, and Islam M. T (2020) Can AI help in screening Viral and COVID-19 pneumonia. IEEE Access 8:132665–132676
Guo H, Wu S, Tian YJ, Zhang J, Liu H (2021) Application of machine learning methods for the prediction of organic solid waste treatment and recycling processes: A review, Bioresource Technology, 319, 2021, 124114, doi:10.1016/j.biortech.2020.124114
Helm JM, Swiergosz AM, Haeberle HS, Karnuta JM, Schaffer JL, Krebs VE, Spitzer AI, Ramkumar PN (2020) Machine learning and artificial intelligence: definitions, applications, and future directions. Curr Rev Musculoskelet Med 13(1):69–76
Kaya A, Keceli S, Catal C, Yalic HY, Temucin T, Tekinerdogan B (2019) Analysis of transfer learning for deep neural network based plant classification models, vol 158. Computers and Electronics in Agriculture, pp 20–29
Kryzhanovsky B, Dunin-Barkowski W, Redko V, Tiumentsev Y (eds) (eds) (2020) Advances in Neural Computation, Machine Learning, and Cognitive Research III. Studies in Computational Intelligence. doi:10.1007/978-3-030-30425-6
Lin K, Zhao K, Zhao Y, Kuo J, Deng H, Cui F, Zhang Z, Zhang M, Zho C, Gao X, Zhou T, Wang T (2022) Toward smarter management and recovery of municipal solid waste: A critical review on deep learning approaches, Journal of Cleaner Production 346, 130943, 2022, doi.org:10.1016/j.jclepro.2022.130943
Maini R, Aggarwal H (2009) Study and comparison of various image edge detection techniques. Int J Image Process 3:1–12
Malik M, Sharma S, Uddin M, Chen C-L, Wu C-M, Soni P, Chaudhary S (2022) Waste Classification for Sustainable Development Using Image Recognition with Deep Learning Neural Network Models. Sustainability 14:7222. https://doi.org/10.3390/su14127222
Mao WL, Chen WC, Wang CT, Lin YH (2021) Recycling waste classification using optimized convolutional neural network. Resources, Conservation and Recycling, 164, p.105132
Mao WL, Chen WC, Fathurahmman HIK, Lin YH (2022) Deep learning networks for real-time regional domestic waste detection, Journal of Cleaner Production, 344, 2022, 131096, doi.org/10.1016/j.jclepro.2022.131096
Marin I, Mladenovi´c S, Gotovac S, Zaharija G (2021) Deep-feature-based approach to marine debris classification. Applied Science 2021, 11, 5644. https://doi.org/10.3390/app11125644
Mazzeo PL, Argentieri A, De Luca F, Spagnolo P, Distante C, Leo M, Carcagni P (2019) Convolutional neural networks for recognition and segmentation of aluminum profiles. Multimodal Senssing: Technologies and Applications. 10.1117/12.2525687
Melinte O, Travediu A, Dumitriu DN (2020) Deep convolutional neural networks object detector for real-time waste identification, Applied Sciences, 2020, 10, 7301; doi:10.3390/app10207301
Memon MA (2010) Integrated solid waste management based on the 3R approach. J Mater Cycles Waste Manage 12:30–40. 10.1007/s10163-009-0274-0
Meng S, Chu WT (2020) February. A study of garbage classification with convolutional neural networks. In 2020 Indo–Taiwan 2nd International Conference on Computing, Analytics and Networks (Indo-Taiwan ICAN) pp. 152–157. IEEE
Mitra A (2020) Detection of waste material using deep learning and image processing, MSc thesis in Computer Science, California State University, San Marcos
Nanez Alonso SL, Reier Forradellas RF, Pi Morell O, Jorge-Vazquez J (2021) Digitalization, circular economy and environmental sustainability: the application of artificial intelligence in the efficient self-management of waste. Sustain 2021 13:2092. 10.3390/su13042092
Niu B, Feng Q, Yang J, Chen B, Gao B, Liu J, Lidan Y, Gong J (2022) Solid waste mapping based on very high resolution remote sensing imagery and a novel deep learning approach. Geocarto Int. 10.1080/10106049.2022.2164361
Owotogbe JS, Ibiyemi TS, Adu BA (2019) A Comprehensive review on various types of noise in image processing. Int J Sci Eng Res 10:388–393
Rajak A, Shazia H, Bushra M (2020) Automatic waste detection by deep learning and disposal system design. J Environ Eng Sci 15:38–44. 10.1680/jenes.19.00023
Shi C, Xia R, Wang L (2020) A novel multi-branch channel expansion network for garbage image classification. Ieee Access 8:154436–154452
Shi C, Tan C, Wang T, Wang LA (2021) Waste Classification Method Based on a Multilayer Hybrid Convolution Neural Network. Applied Science 2021, 11, 8572. https://doi.org/10.3390/app11188572
Shorten C, Khoshgoftaar TM (2019) A survey on image data augmentation for deep learning, Journal of Big Data (2019) 6:60, doi: 10.1186/s40537-019-0197-0
Sothe C, De Almeida CM, Schimalski MB, La Rosa EC, Castro JDB, Feitosa RQ, Dalponte M, Lima CL, Liesenberg V, Miyoshi GT, Tommaselli AMG (2020) Comparative performance of convolutional neural network, weighted and conventional support vector machine and random forest for classifying tree species using hyperspectral and photogrammetric data. GIScience & Remote Sensing 57(3):369–394. 10.1080/15481603.2020.1712102
Srinath KR (2017) Python – The Fastest Growing Programming Language, International Research Journal of Engineering and Technology (IRJET), 4, pp 354–357
Tao J, Gu Y, Hao X, Liang R, Wang B, Cheng Z (2023) Combination of hyperspectral imaging and machine learning models fo fast characterization and classification of municipal solid waste, Resources, Conservation and Recycling, 188(2023) 106731
Togacar M, Ergen B, Comert Z (2020) Waste classification using AutoEncoder network with integrated feature selection method in convolutional neural network models, Measurement, 153, 2020, 107459, doi:10.1016/j.measurement.2019.107459
Vrancken C, Longhurst PJ, Wagland ST (2017) Critical review of real-time methods for solid waste characterization: Informing material recovery and fuel production, Waste Management 61 (2017) 40–57
Wilts H, Garcia BR, Garlito RG, Gómez LS, Prieto EG (2021) Artificial Intelligence in the Sorting of Municipal Waste as an Enabler of the Circular Economy, Resources 2021. 10:28. https://doi.org/10.3390/resources10040028
Wilts CH, Berg H (2018) Digital cycles and digital scrap: How digitization improves resource efficiency in the circular economy; and where it tends to cost resources. J Waste Recycling 3:1–4
Wu TW, Zhang H, Peng W, Lu F, He PJ (2023) Applications of convolutional neural networks for intelligent waste identification and recycling: A review, Resources Conservation and Recycling, 190(2023) 106815, https://doi.org/10.1016/j.resconrec.2022.106813
Xia W, Jiang Y, Chen X, Zhao R (2021) Application of machine learning algorithms in municipal solid waste management: A mini review. Waste Manage Res 40:609–624. 10.1177/0734242X211033716
Yang M, Thung G (2016) Classification of trash for recyclability status, CS229 project report, 2016, pp.388–393

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

Intelligent solid waste classification system using combination of image processing and machine learning models

Status:

Version 1

Abstract

Figures

1. Introduction

2. Methodology

2.1 Solid waste input data sets

2.2 Image pre-processing

2.3 Image Augmentation and Segmentation

2.4 Machine learning models

3. Results And Discussions

3.1 JONET Model

3.4 Multiple Object Model

4. Conclusions And Recommendations

References

Additional Declarations

Status:

Version 1