Progressive feature enhancement network for object detection in low-light conditions

Object detection in low-light environments is a challenging problem in the field of computer vision. Traditional methods optimize image enhancement and object detection separately, which not only leads to overly complex models but also prevents the enhanced features from being effectively utilized for object detection. Enhancing low-light images solely for human visual perception does not always benefit the object detection task. To address this issue, this paper proposes a Progressive Feature Enhancement Network for object detection. During the feature extraction stage, the network employs Multi-Scale Receptive Field Convolution (MSRFConv) to progressively enhance the image multiple times. The multi-scale features obtained from these enhancements are then deeply cross-fused to provide richer features for subsequent object detection. By introducing dynamic convolution technology, a feature fusion module based on dynamic convolution is proposed. This module adaptively adjusts its parameters based on the input feature content, enhancing the model’s generalization capability for complex lighting scenarios. Furthermore, a training-aware dynamic IoU loss function is introduced. This loss function incorporates a dynamic scaling factor correlated with the training epoch to accelerate the model convergence process. Experimental results on the public datasets ExDark and DarkFace demonstrate the effectiveness of the proposed method, showing improvements of 2.6% and 1.3% in the mAP50 metric, respectively.