Zero-reference generative exposure correction and adaptive fusion for low-light image enhancement

Abstract

Existing low-light image enhancement methods have the problem of difficulty in enhancing dark areas while controlling overexposed areas in natural images. To address this issue, a Generative Exposure Correction method based on Retinex theory is proposed in this paper, in which the Pseudo-Exposure Residual map and illumination map are deeply coupled based on the proposed intensity compensation prior to constrain the generative network’s output in order to simultaneously deal with overexposure and underexposure. Furthermore, to enhance the effect and prevent over-correction, an exposure fusion technique is proposed, which adaptively selects the best exposure area from the two corrected images and achieves a globally balanced exposure by using an intensity correction compensation operator. More importantly, our proposed method does not require the collection of additional external datasets, which also overcomes the difficulty of data acquisition. Experimental comparisons of our method with the other seven state-of-the-art methods on five public datasets demonstrate that our method achieves the best performance in terms of detail enhancement and natural color preservation.