YOLO26-USW: A Slice-Aided Semantic–Detail Fusion Detector for Small Insects in High-Resolution Agricultural Light-Trap Images

Abstract

Detecting small insects in high-resolution agricultural light-trap images is difficult because resizing obscures fine morphology and densely distributed insects often appear similar. We developed YOLO26-USW using PVTv2-B1-based UNet-PVT feature enhancement, the Semantic–Detail Fusion Module (SDFM), and Wise-IoU v3 (WIoU v3). Slicing Aided Hyper Inference (SAHI) was evaluated separately on full-resolution images. The dataset contains 2505 original images (2004 training, 250 validation, and 251 test) with 9887 annotated insect instances. Slicing yielded 8939, 979, and 1128 target-containing 640 × 640 patches, comprising 13,162 slice-level annotations. On the pre-generated sliced-image benchmark, YOLO26-USW achieved 90.8% precision, 91.1% recall, 94.5% mAP@0.5, and a 91.0% F1-score, exceeding YOLO26 by 6.4, 5.4, 5.5, and 6.0 percentage points, respectively, at a native throughput of 140.38 FPS. On 251 full-resolution test images, direct inference produced no true-positive matches at IoU = 0.5, whereas SAHI achieved 66.90% precision, 96.15% recall, 78.90% F1-score, and 77.69% mAP@0.5. Processing time increased from 0.0573 s/image (direct) to 3.9366 s/image (SAHI). In selected sparse, dense, clustered, and partly occluded scenes, the proposed method showed clearer localization in some cases, although closely spaced or visually similar insects remained challenging. The results illustrate the gains in small-insect detection alongside the additional cost of full-resolution sliced inference.