Adaptive Loss-Guided Multi-Stage Residual ASPP for Lesion Segmentation and Disease Detection in Cucumber under Complex Backgrounds

Abstract

- In complex environments, the performance of leaf segmentation models for cucumber diseases tends to decrease due to the overlapping of obstructions such as shadows and leaf debris, as well as the impact of uneven illumination. This, in turn, directly affects the subsequent disease detection tasks. Furthermore, the imbalance in pixel ratio between background and lesion areas can seriously impact the accuracy of lesion extraction. To address these issues, we propose an image segmentation framework based on a two-stage Atrous Spatial Pyramid Pooling (ASPP) with adaptive loss for precise cucumber lesion segmentation under complex scenarios. The model architecture guided by adaptive loss can reduce the loss weight generated by pixels that are easy to classify. Therefore, during the model training process, more attention will be paid to pixels that are challenging to classify, thereby improving the accuracy of lesion segmentation. The two-stage model, which we refer to as the LS-ASPP model, consists of Leaf-ASPP and Spot-ASPP. In the first stage of this model, images are processed through the Leaf-ASPP architecture, which extracts leaf contours from the complex background. In this stage, we incorporate attention modules and residual structures into the ASPP framework, creating an improved residual ASPP network. This can capture multi-scale semantic information of diseased leaves, enhancing the model's edge perception capabilities. In the second stage, the Spot-ASPP model segments the lesion areas from the extracted leaf contours. We adjust the dilation rate of the Atrous Spatial Pyramid Pooling (ASPP) to capture smaller targets, and introduce a Convolutional Attention Block Module (CABM) to highlight important information. Compared to existing deep learning models, this framework can improve semantic segmentation accuracy under complex conditions.