Research on Low Illumination Target Detection Method for Prefabricated Beam Field Based on Retinex

There is a problem of low illumination in the video surveillance of the precast beam field, which seriously affects the effect of the intelligent analysis algorithm. Based on the in-depth study of the low-illumination images of beam fields, this paper proposes a decomposition network module based on the encoding-decoding architecture of Retinex theory, which uses the improved U-Net structure to extract features, maps the features to the low-dimensional probability density distribution with the help of normalized streams, and decodes the reflection component features of the input image. The network module is used to replace the feature extraction module of the existing mainstream object detection algorithms, which can significantly improve the recognition effect of the algorithm under low illumination conditions. After embedding the module in the mainstream object detection algorithms YoloV8 and Faster RCNN respectively, the MAP value of object detection remains stable under normal illumination, and the MAP value of object detection is increased by 5% on average under low illumination.