Robust Real-world Image Enhancement Based on Multi-Exposure LDR Images

Robust real-world image enhancement from multi-exposure low dynamic range (LDR) images is a challenging task due to the unexpected inconsistency among the input images, such as the large motion or various exposures. In this paper, we propose a novel end-to-end image enhancement network to solve this problem. After extracting contextual information from the LDR images, we design a novel matching volume to align them by considering the motion and exposure differences among the input images. A stacked hourglass with dilated convolution is further utilized to aggregate the matched feature maps to the final enhanced image. In addition, we design a weakly-supervised pairwise loss function to evaluate the color consistency in the enhanced image, which further boosts the performance. We show the effectiveness of our methods on high dynamic ranging imaging (HDR) and End-to-End image signal processing (E2E-ISP) tasks. Experimental results demonstrate that our model achieves state-of-the-art enhancement performance.