Semantic Segmentation of Rice Leaf Diseases Based on GEA-SegFormer

Rice, as an important crop in China, often suffers serious impacts on its yield due to leaf diseases. Timely and accurate assessment of the degree of rice leaf disease is crucial for the development of effective prevention and control strategies. Semantic segmentation technology can accurately segment the diseased leaf parts, thus assisting farmers in assessing the degree of leaf lesions and taking timely preventive and control measures. However, the complexity of the rice growing environment, the diversity of leaf lesions, and the ambiguity of the edges limit the accuracy of the existing segmentation techniques in leaf disease identification. To address these issues, this paper proposes a GEA-SegFormer rice leaf disease semantic segmentation method based on the SegFormer algorithm. The method is based on SegFormer with Global Attention Mechanism (GAM) and Efficient Channel Attention (ECA).GAM distinguishes the disease from the background more efficiently by enhancing the global context and multi-scale features, while ECA focuses on features related to the shape and texture of the disease spot and improves the edge recognition ability. On the customized dataset, the MIoU, recall and average pixel accuracy of GEA-SegFormer outperform those of SegFormer by 6.7%, 10.3% and 5.2% to 85.8%, 81.2% and 90.3%, respectively, showing higher segmentation accuracy and effectiveness.