Deep learning-based multiscale CNN-based U network model for leaf disease diagnosis and segmentation of lesions in tomato

Abstract

BackgroundThe advent of deep learning technologies has expanded the utilization of convolutional neural networks in diagnosing and segmenting plant leaf diseases. However, the segmentation of lesions, especially those with diverse sizes and shapes, often requires extensive pixel-level annotations. ObjectiveThis study aims to introduce a multiscale U-network for effective concurrent segmentation and diagnosis of tomato leaf lesions. MethodsOur model employs multi-scale residual modules to adapt dynamically to diseases that alter lesion size and shape. The Classifier and Bridge (CB) module is introduced to connect the disease feature extraction stage with the disease spot segmentation stage, providing crucial information about a specific lesion type. During the segmentation stage, the activation map of a specific class is deconvoluted using a combination of up-sampling and convolution. This deconvolution feature is fused with the low-level feature through a skip connection strategy. A limited number of pixel-level labels are used, and the binary cross-entropy loss function is applied for supervised training at each pixel point, directing the feature extraction network to focus on the lesion location. ResultsThe model was evaluated using an original test set and an interference test set simulating noise and light intensity changes. The multiscale U-network demonstrated an average pixel accuracy of approximately 99.2%, a resilience to brightness reduction by about 92.4%, and an ability to handle up to 0.98 salt and pepper noise at a rate of about 99.2%, showcasing its robustness against various image alterations. The GFLOPs for the proposed method was found to be approximately 3.13e+06, indicating its computational efficiency. ConclusionThe experimental results highlight the potential of the proposed multiscale U-network in effectively segmenting and identifying lesions in tomato leaf tissue. These findings suggest promising opportunities for applying this model in the broader field of agricultural disease management.