Slicing-Aided Hyper Inference for Small-Insect Detection and Counting in High-Resolution Agricultural Images: Impact of Framework and Grid Geometry

Source abstract

Small-insect detection and counting in high-resolution agricultural images remain challenging because global resizing suppresses fine-scale target information, while complex crop backgrounds increase visual ambiguity. Slicing Aided Hyper Inference (SAHI) mitigates this limitation through local tile-based inference, whereas Adaptive Slicing Aided Hyper Inference (ASAHI) reduces redundant slicing through resolution-adaptive partitioning. However, the influence of slice-grid geometry on agricultural insect detection remains comparatively underexplored. This study evaluates full-image inference, SAHI, ASAHI, and an aspect-ratio-aware ASAHI adjustment (ASAHI-AR) using YOLOv8n and Faster R-CNN under five-fold cross-validation on INSECT-SOYBEAN-DET-1C, comprising 987 soybean-field images and 6,444 bounding-box annotations. Baseline mAP@[.5:.95] was 0.128 ± 0.009 for YOLOv8n and 0.112 ± 0.012 for Faster R-CNN. SAHI increased recall to 0.674 ± 0.037 and 0.895 ± 0.015, respectively, but increased Faster R-CNN counting RMSE from 3.725 ± 0.357 to 11.209 ± 0.602. ASAHI reduced tile redundancy while retaining detection gains. ASAHI-AR achieved the most favorable overall balance, including the highest YOLOv8n F1-score (0.627 ± 0.020) and Faster R-CNN mAP@[.5:.95] (0.315 ± 0.007). These results show that slicing intensity and grid geometry affect detection–counting trade-offs. The dataset, cross-validation splits, and source code are publicly available.