In the field of computer vision, deep learning has always been a good performance.Target detection method based on deep learning has great potential in infrared image detection and recognition. For the changeable scene of infrared image, large scale variation range and insufficient available features, in this paper, a kind of Dense-yolov5 network structure is proposed, which combines the characteristics of DenseNet network and Yolov5 network, based on the idea of making full use of features to protect the edge of the target. The unit module in Yolov5s is replaced with the user-defined dense connection Denseblock module, and the contrast experiment between Dense-yolov5 network and the original Yolov5 network on the self built infrared data set is carried out. The recall rate, accuracy and mAP value of each target detection are improved by the improved Dense-yolov5 network, and the mAP is improved by 1.92%, recall rate is 1.59% and precision is 0.99%. The detection results show that the network has better recognition effect for infrared targets, especially for small targets with unclear features, and has great value for infrared target recognition and tracking in the future military field.