Aiming at the problems of small and dim target with relatively few pixels, difficult to extract pixel features, low detection probability and high false alarm rate. In this paper, an algorithm based on multi-scale feature enhancement for dim target detection is proposed. Starting from the upper layer of detection task, the image high-resolution reconstruction network based on pyramid architecture is first used to strengthen the small and target region and distribution features, and then the reconstructed target features are sent as learning labels into the Yolo v5s detection framework containing multi-channel convolution kernel for self-supervised learning. Through the case test with different number of target points, it is proved that the mAP value of the enhanced image is more than 4% higher than that of the original image. © 2023 IEEE.