Research on Weakly Supervised Deep Learning Model for Semantic Segmentation of Pest and Disease Regions in Durian Blind Spots

ABSTRACT To address severe canopy occlusion, complex backgrounds, irregular pest‐damage morphology, and the high cost of pixel‐level annotation in durian orchards, this study proposes Dp‐GradCAM, a weakly supervised semantic segmentation method for durian pest and disease damage areas. The method targets typical durian pests and diseases, including leaf blight, leaf spot, algal spot, leaf miner, thrips, and mealybugs, aiming to achieve accurate localization and segmentation using only image‐level labels. Unlike conventional weakly supervised methods that directly rely on coarse activation maps, Dp‐GradCAM first uses a ResNet50 classification network to extract discriminative features from healthy and damaged durian images and then applies Grad‐CAM to generate target‐category heatmaps as initial seed regions. To refine these coarse seeds, ExG filtering is used to remove interference from healthy green leaves and weeds, while DenseCRF optimizes lesion boundaries based on pixel spatial relationships and color similarity. Hole filling and morphological opening are further applied to complete target regions and remove isolated noise, generating high‐quality pseudo‐labels closer to real pest‐damaged areas. Based on these pseudo‐labels, a lightweight DeepLabv3+ segmentation network is constructed by improving the ResNet50 backbone and introducing depthwise separable dilated convolutions, reducing model complexity while preserving multi‐scale semantic representation.