Abstract: In response to the problems of the high missing rate of existing YOLO target detection algorithms in self-built data sets and low detection accuracy caused by complex environments, proposes an improved infrared image recognition algorithm based on YOLO v5. According to the unique properties of infrared data images, the backbone network is redesigned, the OMNI-Dimensional Dynamic Convolution Module (ODConv) is introduced, and the Coordinate Attention (CA) mechanism is improved. In the meanwhile, improve the detection accuracy of small targets and reduce the number of parameters. Secondly, the Decoupled Head (DH) module is introduced to improve the convergence speed of model training. Finally, add the Graph-Shifted Convolution (GSConv) Slim module to reduce the complexity of the model and increase the speed of prediction. The experimental results show that the missed rate of the improved algorithm is reduced by 40.22%, the Floating-Point Operations per Second (FLOPs) is increased by 25% and the average accuracy is increased by 28.32%.