Semi-supervised Zero-Shot Learning in Retinex Decomposition-Based Low-Light Image Enhancement.

Jieyu Pang,Fanqi Yu,Ronggang Wang

DOI: https://doi.org/10.1145/3594315.3594338

2023-01-01

Abstract:The images recorded by people are often subject to multiple degradation challenges due to the shooting environment and shooting equipment. In low-light scenes, images acquired by non-professional equipment are usually accompanied by a large amount of noise, and their contrast and color also have large deviations relative to normal images. Existing low-light enhancement methods are either limited by the number of low-light paired datasets, which tend to overfit the available data, or have difficulty achieving better performance when not using paired data. We propose a Semi-supervised Zero-Shot Learning low-light enhancement method based on Retinex decomposition, which can achieve better decomposition ability on existing paired data with multiple exposure conditions through supervised training, and also improve the overall network generalization using Zero-shot learning method. With a well-designed loss function, our method uses supervised training in the decomposition phase, which makes full use of paired images under different illumination conditions, and Zero-shot training in the illumination enhancement phase, which relies entirely on the unsupervised loss function to guide the enhancement of the illumination components. Our experiments demonstrate that our method has surpassed traditional augmentation methods as well as partially unsupervised methods, and is more generalizable than existing supervised methods.