IMMP-YOLO: A Real-Time Object Detection Algorithm Incorporating Image Enhancement and Pruning Strategy

In real traffic scenarios, vehicle detection under adverse lighting conditions confronts numerous challenges, images usually suffer from low contrast, high noise and color deviation, all of which make object recognition less accurate and slow in real time. Hence this paper proposes a real-time object detection algorithm called IMMP-YOLO that combines image enhancement and pruning strategy. Firstly, the haze and darkness images in the dataset are processed using the de-fogging algorithm RIDCP and the low-light enhancement algorithm GSAD to restore the clarity and contrast of the images in order to improve the quality of the input images. Secondly, FasterNet is introduced for reconstructing the YOLOv7 backbone, which greatly simplifies the original network structure; the neck structure has been enhanced, specifically the original ELAN structure has been substituted with the lightweight RepNCSPELAN4 to augment the feature extraction capacity; and the original detection head is replaced by the decoupled head structure to improve the detection accuracy and robustness. Finally, the L1 pruning strategy is used to compress parameters and GFLOPs of the improved model by 52% and 51%, respectively, by removing the redundant network structure, while the detection accuracy is only lost by 0.77%. Experiments on the adverse lighting vehicle driving dataset indicate that the proposed method outperforms the original YOLOv7 and other mainstream object detection algorithms in detecting accuracy and real-time, with mAP0.5, mAP0.95 reaching 74.29% and 51.81%, improving by 4.79% and 5.52% compared to the baseline, respectively. The detection speed reaches 117.9 FPS, an increase of 54.1 over baseline.