YOLO-CIR: The network based on YOLO and ConvNeXt for infrared object detection

Transfer learning is widely used in infrared object detection algorithms, but these algorithms developed from visible usually ignored the characteristic of infrared images. In this paper, we propose a new object detection algorithm YOLO-CIR based on YOLO and ConvNext for infrared (IR) images. Specifically, to accommodate high-bit-width infrared images, we propose the augmentation algorithm for infrared images and improve the pre-processing algorithm without bit-width compression. In addition, a multi-scale feature extraction network based on ConvNext was built to adapt the infrared images of low-resolution. Moreover, the coordinate attention module is introduced in the ConvNeXt block to focus on targets and suppress the background, and a split attention module is introduced in the neck to enhance feature fusion ability. The algorithm offers significant performance improvements over prevalent object detection algorithms, comparative experiment on the FLIR dataset shows that the algorithm outperforms the YOLOv5 by 3% and Faster R-CNN by 5.6% in map50 and has significant advantages in parameters and FLOPs.