Answers are grounded in this paper and cite their sources.
Mango leaf disease segmentation is an essential foundation for accurate disease diagnosis and intelligent grading. The size and shape of mango leaf diseases vary significantly at different times, making it difficult for mainstream semantic segmentation methods to segment disease areas accurately. Therefore, this paper proposes a method called MAU-Net for fine segmentation of mango leaf diseases over the whole period. The MAU-Net is based on the traditional Unet architecture, integrates the Self-Aligning Attention Feature Fusion (SAFF) module and the Multiscale Feature Enhancement (MFE) module, and designs a new loss function DF_Loss. Specifically, the designed SAFF module changes the traditional Unet's skip-connection approach by fusing the global and local two-branch attention mechanisms. It enhances the attention to crucial leaf and disease features at different levels and thus retains richer semantic information about mango leaf diseases. The designed MFE module aims to solve the problem of complex multi-scale disease segmentation in different periods of mango leaves by introducing different scales of cavity convolution to enhance the extraction of disease features at different scales. The designed DF_Loss combines the idea of the similarity measure in Dice Loss and the advantages of the attentional conditioning mechanism in Focal Loss with an additional conditioning factor. It allows the model to focus more on pixels that are difficult to categorize during the learning process, thus improving the segmentation accuracy. MAU-Net achieved 99.21%, 84.33%, 97.1%, and 96.94% of leaf IoU, disease IoU, F1, and mPA metrics on the mango leaf disease dataset. It improved 0.36%, 4.88%, 3.9%, and 1.91% over UNet, and 5.59%, 0.19%, 1.6%, and 2.26% over DeepLabv3+, respectively. Therefore, the present study may provide an accurate method for segmenting mango leaf spots over the whole period and provide a sufficient basis for the accurate analysis of mango leaf diseases.
No takes yet. Share an insight, caveat, or question.
Tang et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: