In response to the problems of low image information entropy and insufficient recognition accuracy in existing recognition methods， this article adopts an improved YOLOv5 algorithm for low light image recognition. By enhancing the brightness of local dark areas， improving the overall brightness of the image， and aggregating multi-scale features， a low light target recognition architecture is constructed. This architecture can utilize convolutional neural networks to optimize the scale of the target region of interest （ROI） to ensure that image details are not lost. The experimental results showed that based on 5 000 samples， the information entropy value of the improved method increased to 6.5， and the recognition accuracy reached 98%， significantly better than the 68%~72% of the control group. This method achieves precise low light image recognition， effectively improving image visibility and providing an effective solution for image recognition tasks in low light environments.