MRT-YOLO: A Fine-Grained Feature-Based Method for Object Detection

Object detection is an essential component of autonomous driving, unmanned aerial vehicle (UAV) reconnaissance, and other domains. It equips drones and vehicles with the capability to perceive and comprehend their surrounding environment, making it a crucial technology for achieving safe and reliable autonomous driving as well as UAV spot reconnaissance. This paper proposes an end-to-end, high-precision, multi-scale real-time detection algorithm called MRT-YOLO based on YOLOv8. Firstly, in the feature downsampling process of the backbone network, we extend the channel depth to enhance the model’s learning capability for fine-grained features and thereby improve its performance in retaining feature information. Secondly, we enhance the cross-stage partial layer version 2 (C2f) module in YOLOv8 by incorporating a channel self-attention mechanism within it, which optimizes performance through effective feature interaction and integration. Simultaneously, we also employ an improved bidirectional feature pyramid network (BiFPN) and introduce the proposed multi-scale feature learning (MFL) module to further enhance the model’s feature extraction ability. In this study, we fuse the feature maps (C2, C3, C4, and C5) from the backbone network to generate a new feature map C6, thus increasing cross-connections between low-level and high-level features. Lastly, a multi-scale small object detection structure is designed to enhance recognition sensitivity toward densely distributed small objects. The proposed algorithm’s effectiveness and superiority are demonstrated through experiments conducted on two datasets: VisDrone (UAV vision dataset) and BDD100K (automatic driving dataset).