A Low-Light Image Enhancement Algorithm Incorporating Cross-Mixed Attention and Receptive Field Expansion Mechanism

Abstract

Aiming at the conventional low-light enhancement algorithms for low-light image enhancement with problems of loss of details, low contrast and low color saturation, a detection algorithm called CE-Retinex(Cross Expansion Retinex) by incorporating the cross-mixed attention mechanism and the receptive field expansion mechanism is proposed and demonstrated for the first time. It mainly consists of three parts: initial module, optimization module, and detail restoration. Firstly, in the initial module, the image is decomposed into reflectance and illuminance using two different U-shaped networks with multiple layers of convolution. Secondly, in the optimization network module, the image brightness is enhanced using multiscale lighting mechanism. Denoising is also performed. After that, a multi-layer convolutional fusion crossmixed attention mechanism is used to filter the information from the four dimensions of channel, spatial, vertical and horizontal attention, so that it can be effectively reduce the negative effects of the low-light image enhancement process. In the detail restoration module, the receptive field expansion mechanism is utilized to enhance the receptive field and strengthen the detail information. Also, the color consistency loss function is used to recover colors in the loss function. The CE-Retinex algorithm was experimentally analyzed on the LOL dataset, and the PSNR of the CE-Retinex algorithm was 25.33, and the NIQE was 3.37, and the subjective feelings and objective evaluation indicators have effectively improved. So the proposed algorithm could be effectively solve the problems of loss of detail, low contrast, and low color saturation in low-light image enhancement.