RS-YOLO: A highly accurate real-time detection model for small-target pest

Abstract

Pest detection in agriculture is critical for effective pest control in farming. However, existing algorithms for real-time small-target pest detection often struggle to balance detection speed and accuracy. The YOLOv8 model has been widely adopted in various fields of agricultural production due to its robust performance across diverse scenarios and datasets. Therefore, this paper proposes RS-YOLO, a pest detection model optimized for dense, small-target pests, based on the YOLOv8n model. To enhance the ability of the model to capture detailed features, the Parallel Bottleneck(PBN) module is proposed by adding parallel structures to the original bottleneck module. To mitigate feature loss caused by repeated downsampling, the efficient multi-scale attention(EMA) module introduced into the spatial pyramid pooling fast(SPPF) module. Additionally, the up-sampling method is modified to bilinear interpolation upsampling(BLIU) to improve the clarity of feature maps during upsampling process. These improvements have greatly enhance the ability of the model to retain small target features as well as detailed information, boosting both accuracy and real-time performance of several classical crop pest detection. The experiments were conducted using the pest dataset collected by the automatic pest information collection system. The results showed that the mAP@.5 and mAP@.5:.95 of RS-YOLO reached 96.6% and 81.7%, representing 2.8% and 6.8% improvement over the baseline model, respectively. Compared to mainstream detection models such as Faster R-CNN, YOLOv5s, YOLOv7-tiny, YOLOv8-p2, YOLOv9s, and YOLOv10s, as well as the self-developed model YOLOv5-GCD, our model has demonstrated superior performance.