Research On Rice Disease Recognition Method Based On Deep Residual Shrinkage Networ

As one of the world’s major bulk crops,China’s rice production reaches 210 million tons in2021,and the timely and effective diagnosis and treatment of rice diseases is of great significance to improve yield and quality.Early methods of disease identification such as manual visual inspection and pathology detection in the laboratory can hardly circumvent the problems of difficult feature extraction,lack of generalization ability,and difficulty in large area promotion.Deep learning techniques can achieve automatic extraction of disease features for crop disease identification without manual labeling of samples.To break through the various limitations in rice disease recognition tasks,this paper conducts a study on rice disease recognition based on deep residual shrinkage networks.Based on the deep residual shrinkage network,feature fusion and CBAM(Convolutional Block Attention Module)attention mechanism are introduced to construct a rice disease recognition model ICDRSN(InceptionA and CBAM based DRSN),which realizes multi-scale feature mining under multi-dimensional receptive fields and effective extraction of features in spatial and channel dimensions.The network performance under different module structure combinations is compared and analyzed,and the optimal parameter combination is obtained to improve the classification accuracy of the model.The model is trained using the self-built rice disease data set.The recognition accuracy on the test set is 98.65 %,which verifies the effectiveness of the model.In order to improve the speed of network recognition,based on the ICDRSN network model,the deep separation convolution structure and SimAM(Simple,Parameter-Free Attention Module)parameter-free attention mechanism are further integrated to construct DSDRSN.(Depthwise Separable Convolution and SimAM based DRSN)rice disease identification model.The deep separable convolution is used for split feature extraction,and the network parameters and computation are reduced by combining the parameter-free attention mechanism to maximize the learning of the protruding neurons.The model was trained using the same rice disease data set,and the average recognition accuracy reached 99.39 %.The recognition speed and average recognition accuracy were 8.96 and 0.74 points higher than the ICDRSN recognition results,respectively,which verified that the model was superior to ICDRSN.A rice disease recognition system based on residual shrinkage network is designed and implemented to realize intelligent recognition of rice disease images.Based on the two network models,the results show that the system recognition performance under the embedded DSDRSN model is better.