Research On Fine-grained Image Recognition Of Crop Diseases Based On Convolutional Neural Network

Crop diseases have an important impact on agricultural production and economic efficiency,and timely detection of crop diseases and accurate determination of disease species are important for both crop safety and control of disease spread.The traditional naked-eye observation method requires professional knowledge and experience of those engaged in agriculture,and the quality of diagnosis depends entirely on their professional level and experience,and accuracy and timeliness are difficult to guarantee.Most of the relevant research in recent years has been to build modern automatic recognition systems applied to crop disease recognition tasks.Among them,the research on crop disease recognition based on deep learning has achieved more advanced research results,but the application capability in the actual crop growth environment is poor.Compared with crop disease images in a controlled laboratory environment,the actual growth environment is complex and susceptible to external disturbing factors.Therefore,there is a current need to further study automated disease recognition under the actual growth environment of crops to improve the practical application value of related research.In this paper,we take crop disease image datasets in a controlled laboratory environment and crop growth environment as research objects,propose a fine-grained image recognition model for diseases in a crop growth environment,and study the application of fine-grained image recognition methods based on feature coding on crop disease recognition tasks.The main contents are as follows.(1)In this paper,we propose a dual-label crop and disease classification method that separates crop and disease recognition and classifies them independently,reducing the interference between crop and disease features and irrelevant information.The traditional "crop-disease" pair(i.e.,the association between a crop and a disease)single-label classification method for crop disease identification leads to the extraction of irrelevant feature information by the model,which interferes with the accuracy of crop disease classification by the model.We combined crop and disease dual-label classification methods and constructed a finegrained crop disease recognition model based on a bilinear pooling approach to independently classify crops and diseases so that the model can focus more on the distinguishing and detailed features in crop disease images.Meanwhile,this paper uses a transfer learning technique to transfer model parameters trained on a crop disease image dataset in the controlled laboratory environment to a crop disease identification model in a crop growth environment,which improves the application capability of the model in a crop growth environment.(2)We further investigate the application of models using different bilinear pooling methods on crop and disease multi-task recognition.Feature interaction and feature learning between different layers in a convolutional neural network are interrelated and mutually reinforcing,and cross-layer interaction can capture the feature relationships between the layers and improve the representation capability of the network.We proposed a dual-stream hierarchical bilinear pooling model to improve the application of the model in crop growth environments by fusing feature information between different layers through a hierarchical bilinear pooling framework.Meanwhile,the optimal weights between two classification tasks of crops and diseases are explored by using homoscedastic uncertainty to automatically learn the task weights,thus allowing the model to achieve better classification performance in both crops and diseases.It further improves the application value of related research in real-world environments and contributes to multi-tasking research on crop and disease recognition.