Design Compact YOLO based Network for Small Target Detection on Infrared Image

most YOLO object detection neural networks prefer to focus on traditional RGB image, but previous studies rarely consider special YOLO network with compact architecture for infrared image. In this paper, we analyze original YOLO network architecture, and we propose a compact YOLO based network by using different blocks from previous work for small target detection on infrared image. We use small target detection layer, GhostConv convolution, Focus structure, Focal EIOU Loss and soft NMS modules to improve original YOLO network structure, which improves the accuracy and speed of target detection for infrared image. The experimental results show that the accuracy of target detection of original model can be effectively improved by adding the small target detection layer, the Focus structure and the BI-FPN structure. By replacing the GhostConv convolution, the calculation speed is significantly improved. At last, after the model is quantized and deployed with NCNN framework, the object detection speed can reached to 35ms per infrared image.