Abstract: With the development of deep learning technology,convolutional neural network has been the mainstream method for plant leaf disease recognition and disease spot segmentation.Aiming at the problems of different sizes and irregular shapes of tomato leaf lesions,need for a large number of pixel-level labels,a novel multi-scale U network is proposed,which realizes tomato leaf lesion segmentation and disease recognition simultaneously.For disease feature extraction,a multi-scale residual module including different sizes of receptive fields is used to extract disease features according to the different disease spot size and shape.The CB module (Classifier and Bridge) is introduced to connect the disease feature extraction stage with the lesion segmentation stage,which classifies the disease and also reversely generates an activation map of specific class according to the classification result.This activation map contains the specific type of lesions label information.In the segmentation stage,upsampling and convolution are used to deconvolve the activation map.The deconvolution feature and the low-level feature are merged by the jump connection method.In order to make lesion location segmentation more accurate,a few of pixel-level labels are used for training to minimize two-class cross-entropy loss.In the experiment,the original samples and samples with simulated noise and light intensity are used to verify the performance of disease spot segmentation and disease recognition of our method.On the original sample set,the average pixel accuracy,average intersection ratio,and frequency weight intersection ratio of our method reaches 86.15%,75.25%,and 90.27%,respectively.In the interference sample with 30% brightness reduction,salt and pepper noise,the recognition accuracy of our method obtains 95.10% and 99.20% respectively.Experimental results show that the proposed method can achieve improvement in segmentation and recognition of tomato leaf lesions simultaneously.

Key words: Convolutional neural network, Image recognition, Image segmentation, Leaf diseases, Multiscale