LLEYOLO: A target detection algorithm based on improved YOLOv5 for low-light environments

The diversity of low-light environments poses severe challenges to various object detection algorithms, limiting their application in complex scenarios such as autonomous driving. Existing research mainly focuses on "enhance first, then detect" to achieve target detection in low-light scenes. However, these methods can only improve the perceived quality of human vision and are not designed for specific low-light tasks. This paper proposes a lightweight low-light image object detection method LLEYOLO. First, we design a lightweight Attention (EPCA) model for the problem that low-light image features are difficult to extract. This module distinguishes the intensity of feature extraction according to the light intensity of the input image, to achieve the selective expression of low-light feature maps. Second, to address the feature loss problem caused by pooling, we construct a new deformable spatial pyramid pooling (DSPP). DSPP breaks the fixed sampling position of conventional convolution, adjusts some sampling points through changes in the lighting environment, and assigns a weight value to each sampling point for screening. Finally, considering that the category information between objects under low illumination is not obvious enough, the accuracy of the prediction frame is not enough. We propose RIoU as a loss function for bounding boxes so that the network can be optimized towards a high overlap with ground truth boxes. Experiments show that the LLEYOLO model reduces missed and false detection rates on low-light datasets. Compared with the classic YOLOv5 algorithm, it has a significant improvement in location information and category accuracy.