RUCARDAS: Retinex-Inspired Unrolling with Channel Attention Residual Dense Cooperative Prior Architecture Search for Low-Light Image Enhancement

In order to study the image enhancement problem in terms of image color, noise and detail issues, this study proposes a Retinex-inspired Unrolling with Channel Attention Residual Dense Cooperative Prior Architecture Search (RUCARDAS) based on the lightweight architecture RUAS, which introduces a channel attention residual dense convolution operation in the search space, and improves the loss function. First, the method utilizes the basic principle of Retinex to generate light estimation module and noise removal module. It also introduces a channel attention residual dense convolution operation in the subsequent knowledge distillation search space, and searches for lightweight architectures using the improved search space; finally, it improves the loss function of the referenceless learning by combining multiple loss functions and reasonably assigning weight coefficients, so that the enhanced architecture can be searched for by the Retinex-inspired cooperative a priori search network. Weighting coefficients, so that the enhanced image restoration effect is better. The experimental results show that the RUCARDAS model proposed in this paper can effectively improve image reproduction, brightness, contrast and effectively suppress noise while maintaining the advantages of the lightweight framework. In terms of subjective experimental viewing effect and objective evaluation index, RUCARDAS has a good performance effect, and part of the performance is better than the original model and mainstream algorithms.