YOLO-ASPE: ENHANCING SMALL PEST DETECTION IN AGRICULTURAL IMAGERY THROUGH AXIS-SEPARATED POSITIONAL ENCODING

Detecting small pests in agricultural imagery is challenging because insects often blend into foliage and occupy only a few pixels in high-resolution photographs. We propose YOLO-ASPE, a lightweight yet effective modification of YOLOv8 that introduces novel C2f-ASPE (Axis-Separated Positional Encoding) blocks that embed directional spatial awareness into the backbone.Conventional attention modules compress spatial dimensions into scalar representations, eliminating the precise location data essential for detecting minute targets. Our architecture addresses this limitation through independent axis-wise feature encoding that maintains positional awareness across both image dimensions. Trained and evaluated on the AgroPest-12 dataset (Majumdar, 2025), a publicly available collection of 13,143 annotated images spanning12 agricultural pest classes, YOLO-ASPE achieves 84.3% mAP@0.5, outperforming the YOLOv8s baseline by 4.7 percentage points. Improvements are particularly pronounced for small objects (+8.5 points) and morphologically similar species prone to inter-class confusion (31.6% reduction in misclassification). Despite the added attention modules, the model remains efficient, running at 48 FPS on an NVIDIA Jetson Orin, with only 12.8M parameters suitable for real-time field deployment. Our modification is deliberately minimal and restricted to the backbone; the neck and detection head remain unchanged