Semantic Segmentation of Cucumber Leaf Disease Spots Based on ECA-SegFormer

Abstract

Accurate semantic segmentation of disease spots is critical in the evaluation and treatment of cucumber leaf damage. To solve the problem of poor segmentation accuracy caused by the imbalanced feature fusion of SegFormer, the Efficient Channel Attention SegFormer (ECA-SegFormer) is proposed to handle the semantic segmentation of cucumber leaf disease spots under natural acquisition conditions. First, the decoder of SegFormer is modified by inserting the Efficient Channel Attention and adopting the Feature Pyramid Network to increase the scale robustness of the feature representation. Then, a cucumber leaf disease dataset is built with 1558 images collected from the outdoor experimental vegetable base, including downy mildew, powdery mildew, target leaf spot, and angular leaf spot. Tested on the dataset, the Mean Pixel Accuracy of ECA-SegFormer is 38.03%, and the mean Intersection over Union is 60.86%, which is 14.55% and 1.47% higher than SegFormer, respectively. These findings demonstrate the superiority of ECA-SegFormer over the original SegFormer, offering enhanced suitability for precise segmentation of cucumber leaf disease spots in the natural environment.