Boundary-Refined DeepLabV3+ for Crop Disease Detection in Greenhouse Vegetable Images

Abstract

Accurate pixel-level delineation of crop disease in greenhouse images is challenging due to weak lesion margins, scale variations, leaf-vein interference, specular highlights and partial occlusion. This paper introduces a boundary-refined DeepLabV3+ model that keeps the encoder-decoder efficiency of the original framework and adds a boundary supervision branch, uncertainty-aware cross-scale fusion, and an adaptive refinement gate. The network is tested on a set of 6,840 curated greenhouse images that include leaves of tomatoes, cucumbers, peppers and eggplants, as well as 8 disease categories and healthy tissue. Under a fixed split by greenhouse compartment, the proposed model has achieved a mean intersection over union of 89.7%, a mean F1 score of 94.8%, and a boundary F1 of 86.9% at a two-pixel tolerance. The corresponding values are 3.8, 2.4 and 7.6 percentage points higher than those of the standard DeepLabV3+. The mean Intersection over Union (IoU) under low-illumination and condensation-blur conditions are 3.1 and 2.7, respectively. Ablation studies show that boundary supervision is responsible for most of the contour improvement, and uncertainty-aware fusion reduces false lesion expansion along veins. The model has 31.6 million parameters and, after mixed-precision optimisation, runs at 18.7 frames per second on an embedded graphics chip. Based on the above results, explicit boundary reasoning can improve the precision of disease-area estimation without sacrificing the efficiency required in practice; it is thus suitable for greenhouse scouting, targeted spraying and longitudinal severity assessment.