A Attention-Enhanced Lightweight Object Detection for Rice Pest Identification Using YOLOv8n with CBAM and BiFPN

Abstract

The agricultural crop of rice supports the food security of over half of the world but the infestations of pests are considered to be one of the major causes of the loss in yield with the worst line of loss up to 80 percent. The original method of detection, manual scouting, is subjective, time-intensive, and can hardly be applied to large farms. In this paper, a lightweight, real-time object detection model that identifies pests on the rice will be presented based on the addition of Convolutional Block Attention Module (CBAM) and Bidirectional Feature Pyramid Network (BiFPN) neck to the YOLOv8n framework. On a single 26-class rice pest dataset of 11,319 images collected under four Roboflow sources: YOLOv8n (baseline), YOLOv8n+CBAM, the proposed YOLOv8n+CBAM+BiFPN, RT-DETR, Faster R-CNN,

and Florence-2 in zero-shot mode, we compare six detection methods under the same conditions. The proposed model has a precision of 0.5888, a recall of 0.4957, mAP50 of 0.4694 and mAP5095 of 0.3143 at a

constant inference latency of 2.40 ms per image, more than 60.3 times faster than the YOLOv8n baseline with an almost identical mAP50 gap We also demonstrate that the depthwise separable convolutions of BiFPN counterintuitive reduce the inference latency below the baseline, and zero-shot inference on Florence-2 does not work at all without domain adaptation.