Small-Object-Enhanced YOLOv8 for Crop Disease Detection in Greenhouse Vegetable Images

Abstract

Automated disease inspection of greenhouse vegetables is less limited by the visibility of heavily infected leaves and more so by the reliable location of early lesions covering only a few dozen pixels. SE-YOLOv8 is a small-object-enhanced detector in this study that maintains a high-resolution P2 path, performs adaptive bidirectional cross-scale fusion, adds a lightweight coordinate attention module, and uses a scale-balanced localization objective. A greenhouse dataset with 12,480 images of tomato, cucumber, pepper and lettuce was divided into five disease categories plus a healthy background, and 31,762 annotated lesions were included. Under a fixed 640 × 640 input, the proposed model achieved 94.8% precision, 92.6% recall, 95.7% , and 72.4% . Relative to YOLOv8s, recall for lesions smaller than 32 × 32 pixels increased by 8.9 percentage points while parameters rose from 11.2 M to 12.7 M. The model sustained 48.6 frames/s on an NVIDIA Jetson Orin NX and reduced the expected calibration error from 6.8% to 3.9%. Ablation and robustness experiments show that the high-resolution branch contributes the most to the gain, and adaptive fusion and coordinate attention improve discrimination under glare, leaf overlap and clutter. Therefore, by maintaining fine spatial information and regulating cross-scale flow of information, early disease localisation can be achieved without reducing the throughput of practical greenhouses.