Maize is a major crop in China, with the largest planting area and yield, and also plays an important role in light industry, animal husbandry, and the national economy. Maize diseases not only reduce the maize yield but also affect the development of related industries and economies. At present, the manual method is the main method to identify maize diseases in China. The labor process of using manpower to identify maize diseases is not only inefficient, but also easy to be disturbed by subjective factors such as fatigue and emotion, and can only be identified when the obvious symptoms appear (Chen et al., 2020). Therefore, how recognizing the diseases of maize leaves quickly and accurately and taking appropriate control measures is of great significance to ensure maize production.
The research on crop image disease recognition abroad began in the 1980s. Researchers have extensively used a variety of traditional machine learning methods to study the image recognition technology of agricultural diseases, including the support vector machine classifier method(Semary et al., 2015), PNN method(Shi et al., 2015), K-nearest neighbor classification method(S. W. Zhang et al., 2015), BP network method(Wang et al., 2012), and so on, which has played a positive role in promoting the application of information technology in agricultural disease image recognition research. However, the traditional machine learning method has some shortcomings, such as limited learning and expression ability, manual extraction of features, and unsuitable for processing large amounts of data.
The deep learning method can effectively solve the problem of big data learning and modeling. In recent years, researchers have carried out a lot of research work in agricultural disease image recognition based on deep learning. Hammad Saleem et al. (2020) proposed an image-based deep learning meta-structure model to identify plant diseases. Long et al. (2018) proposed a recognition method based on a convolutional neural network and transfer learning for Camellia oleifera disease image recognition, and the average recognition accuracy reached 96.53%. Based on the characteristics of maize foliar diseases, Zhao et al. (2009) applied the threshold method, area marker method, and Freeman link code method to diagnose five major diseases of maize foliage with an accuracy of more than 80% Y Liu et al. (2018). applied the Triplet loss double convolution neural network structure to study the features of corn images, and then used the SIFT algorithm to extract texture features, and the classification accuracy was above 90%. Zeng and Li (2020)proposed the Self-Attention Convolutional Neural Network(SACNN)to identify crop diseases, and extensive experimental results showed that the recognition accuracy of SACNN on AES-CD9214 and MK-D2 was 95.33% and 98.0%, respectively. Compared with the traditional machine learning methods, a deep learning framework can automatically learn the features contained in the image data. When the data set reaches a certain size, it can achieve better accuracy and robustness in the agricultural disease image recognition task. However, the application of deep learning in agricultural disease image recognition still has some problems, such as large training data set, over-reliance on data annotation, limited generalization ability of the model, and high requirements on hardware computing power.
The deep transfer learning method can use the learned knowledge in the field of big data to assist in the building data model in the field of smaller goals, directly reducing the size of the target domain modeling for data requirements, which includes the research field of agricultural disease image recognition. Researchers have carried out some related research work(Fang et al., 2017; Yuan et al., 2018; K. Zhang et al., 2019), using some existing large image datasets to assist in establishing the image recognition model of target disease with small sample data, and achieved certain results.
Although deep learning models for agricultural disease recognition are becoming more and more mature and some research results have been achieved, however, most of the research is based on disease images collected in the laboratory environment, and few studies focused on disease recognition in the actual farmland environment. When these methods are applied to the actual farmland environment, the detection and recognition results are easily affected by the complex environment and the image shooting environment. The recognition accuracy will be greatly reduced, and the applicability is poor with limitations. Compared with the existing methods, the main contributions of this paper are as follows:
We proposed an effective maize disease identification method in complex environments based on cascade networks and two-stage transfer learning. The cascade networks were composed of a Faster R-CNN leaf detector (denoted as LS-RCNN) and a CNN disease classifier (denoted as CENet).
Two-stage transfer learning strategy was proposed to successfully perform transfer learning to train the disease classifier CENet. The transfer of the pre-trained model allowed the model under training to converge faster and to recognize image features with higher accuracy.
We constructed a maize disease data set containing 7144 images, including 3563 images in the natural environment with a more complex background, and 3581 images in the laboratory environment
Fifteen augmentation methods were performed on the existing image data (especially the natural environment) for data enhancement to achieve the purpose of increasing data volume, enriching data diversity, and improving the generalization ability of the model. Also, we investigated the effects of different numbers of amplified images and different amplification methods on the recognition performance.