Abstract:According to the characteristics of infrared images, a YOLOv5-IF method is proposed. The feature extraction network based on the residual mechanism is used to realize the efficient interaction of information between different feature layers and obtain richer target semantic information. By improving the detection scheme of YOLOv5 and adding a larger-scale detection head, the detection probability of small and medium targets in infrared images is effectively improved. Aiming at the problems of limited computing platform resources and real-time demand, the Detection Block module is designed, and the feature integration network is constructed. This module can not only improve the detection accuracy of the method, but also effectively reduce the number of model parameters. On FLIR infrared automatic driving data set, the average accuracy of the proposed method is 74 %, and the parameter is only 19.5 MB, which is better than the existing methods.