Small object detection algorithm based on improved YOLOv5s
Qinggui Huang, Cuihua Tian, Xiangcheng Lv, Chaoxu Lin · 2024
As a technology that has developed over 20 years, object detection has matured significantly with the emergence of practical methods such as Faster R-CNN, RetinaNet, and YOLO, which are widely adopted in industry [1]. However, the challenge of poor performance in detecting small objects remains unresolved. Small objects are defined as those occupying a small fraction of pixels in the input to a computer. Due to the binary nature of digital processing, features of small objects are often filtered out during feature extraction, demanding high precision and accuracy from detection algorithms. Addressing issues like missed detections and false alarms that mainstream object detection algorithms may encounter when handling distant small objects in images, we propose an enhanced YOLOv5s algorithm tailored for small object detection. During model training, we introduce the Focal-EIOU localization loss function to enhance bounding box localization accuracy. In the backbone architecture, a small object detection layer is incorporated to enhance feature collection for small objects. Additionally, in the Neck structure, we utilize the Dy_Sample up-sampling with CBAM structure. Experimental results demonstrate that our enhanced algorithm achieves a 10.4% improvement in [email protected] compared to the YOLOv5s algorithm on the VisDrone dataset, achieving 44.0% [email protected]. Moreover, individual class APs show improvements, and the detection speed meets real-time requirements, making it effective for small object detection tasks.