Image enhancement with bi-directional normalization and color attention-guided generative adversarial networks

The image enhancement aims to improve the quality of images from contrast, detail, and color perspectives by adjusting the color of an image to match the distribution of the high-quality domain. Since the images captured by portable devices often suffer from noise and color bias, this paper designed a novel bi-directional normalization and color attention-guided generative adversarial network (BNCAGAN) for unsupervised image enhancement. Specifically, the bi-directional normalization generator is built upon a feature encoder, an auxiliary attention classifier (AAC), a bi-directional normalization residual (BNR) module, and a feature fusion decoder. With the aid of the AAC and BNR modules, the generator can flexibly control the global style, local details, and color transformation constraints from low-quality and high-quality domains. To improve the visual effect, a spatial color correction module is proposed to assist the multi-scale discriminator in focusing on color fidelity. In addition, a mixed loss, including a content retention loss and an identity fidelity loss, can maintain the structural features to fit the high-quality domain distribution. Extensive experiments on the MIT-Adobe FiveK dataset and DSLR photograph enhancement dataset exhibit that our BNCAGAN outperforms existing methods, and it can improve both the authenticity and naturalness of low-quality images and thus can be widely used for image retrieval preprocessing to improve the understanding of image semantics. The source code is available at https://github.com/SWU-CS-MediaLab/BNCAGAN.