YOLOv8-IRD: Infrared Road Small Object Detection Algorithm Based on Improved YOLOv8

Aiming at the problem that existing model had multiple parameters while showed unsatisfactory recognition accuracy and inaccurate localization in small objects recognition task on infrared roads, this paper proposes a improved YOLOv8-based model for detecting infrared small road objects: YOLOv8-IRD (IRD, Infrared Road Detection). Firstly, to improve the performance of model in object detection while reducing computational complexity, we remove the detection head of size 20×20 and adding a detection head of size 160×160. Secondly, we introduce the Efficient Multi-scale Attention (EMA) for improving the object detection head, addressing the issue of poor localization of small objects. Thirdly, we leverage the RepViT block (RVB) to improve the backbone network, ensuring both lightweight characteristics and robust feature extraction capabilities. Finally, we employ the lightweight upsample module Dysample to alleviate the problem of detail loss during feature map upsampling. The proposed YOLOv8-IRD is evaluated on the FLIR dataset, compared with YOLOv8s and other advanced methods. Experiments show that our approach improves mAP50 by 4.43% and mAP50:95 by 3.30%. Compared to baseline YOLOv8s, the model parameters of YOLOv8-IRD decrease by 46.1%.