A plug-and-play image enhancement model for end-to-end object detection in low-light condition

Although object detection algorithms based on deep learning have been widely used in many scenarios, they face challenges under some degraded conditions, such as low-light. A conventional solution is that image enhancement approaches are used as a separate pre-processing module to improve the quality of degraded image. However, this two-step approach makes it difficult to unify the goals of enhancement and detection, that is, low-light enhancement operations are not always helpful for subsequent object detection. Recently, some works try to integrate enhancement and detection in an end-to-end network, but still suffer from complex network structure, training convergence problem and demanding reference images. To address above problems, a plug-and-play image enhancement model is proposed in this paper, namely, low-light image enhancement (LLIE) model, which can be easily embedded into some off-the-shelf object detection methods in an end-to-end manner. LLIE is composed of a parameter estimation module and image processing module. The former learns to regress lighting enhancement parameters according to the feedback of detection network, and the latter enhances degraded image adaptively to promote subsequent detection model under low-light condition. Extensive object detection experiments on several low-light image data sets show that the performance of detector is significantly improved when LLIE is integrated.