Research on Object Detection Methods in Low-Light Conditions

Feifan Wang, Xi'ai Chen, Xudong Wang, Weihong Ren, Yandong Tang

Abstract: Low-light images are images taken in poorly illuminated environments. Such images suffer from colour distortion, loss of detail and blurriness, which seriously affects the detection accuracy of object detection tasks. In order to improve the accuracy of object detection in low-light images, we propose a low-light image object detection algorithm based on image enhancement. The algorithm is jointly trained on the input side of the YOLOv5 network in combination with an unsupervised low-light enhancement model. The training phase optimises the overall network with the loss of object detection so that the image enhancement results are more favourable for improving the object detection accuracy. In the feature extraction phase, we design a feature enhancement model based on an attention mechanism. Our algorithm is tested on the publicly available ExDark dataset and achieves a mean average precision (mAP) of 79.15%, which is a 4.25% improvement over the baseline.