Exposure Calibration Network With Graph Convolution for Low-Light Image Enhancement

Low-light images are those captured in dim environments with minimal illumination. Such images often exhibit noise, color degradation, uneven exposure, and other undesirable characteristics, making them difficult for human eyes to perceive. Digital camera sensors often struggle to capture vivid images in such circumstances. They cannot autonomously adjust exposure settings for every type of exposure scenario. The image may still have uneven exposure levels even when exposure settings are manually optimized. This necessitates the use of computer vision techniques to calibrate exposure accurately. This article proposes a method based on calibrating the exposure maps along with a graph convolution network (GCN) for low-light image enhancement (LLIE). The model consists of three individual networks that aim to achieve improved performance. The first network, the exposure estimation network, helps to determine the exposure feature maps of the image. The second network, an exposure calibration network, assists in the adaptive calibration of the feature channels. The third network, enhancement block, employs a lightweight graph convolutional network to extract nonlocal features and implement a unique pixel shuffle mechanism instead of basic pooling layers. Finally, features from the previous blocks are added to produce the final enhanced feature map. Extensive simulations showcase our approach’s superior performance.