Tomato leaf disease identification based on attention mechanism and bilinear pool

Abstract

Aiming at the low recognition rate of current crop diseases by convolutional neural network, we propose the network In-SE-Bilinear model to recognize the tomato leaf diseases. In order to reduce the original number of network layers, GoogLeNet network is used as a back-bone. In the process of feature extraction, a Squeeze-and-Excitation (SE) block is connected after each Inception block, and a bilinear pooling layer is added before the connection layer, in front of which the convolution is used to reduce high-dimensional feature information. It is found that the accuracy rate is about 97% for identifying 16 disease classes, which is significantly better than the traditional convolutional neural network, and superior to Inception model, SE model, and bilinear model. Further, some evaluation indicators are calculated by the confusion matrix using test data set, and the results show that the specificity, recall rate and MIoU are 92.52%, 92.44% and 92.55%, respectively. The In-SE-Bilinear method can effectively improve the ability to recognize and classify tomato diseases.