Attention-guided Multiple Receptive Field Residual Block for Single Low-light Image Enhancement

Abstract

Limited by insufficient illumination, the images collected by imaging equipment often have low brightness, low contrast, low signal-to-noise ratio. It severely restricts the development of advanced vision tasks such as target detection and semantic segmentation. To improve the visibility of low-light images, we propose an attention-oriented multiple receptive field residual block (AMRR). Specifically, AMRR first extracts the information of different receptive fields by convolution with varying kernel sizes and then activates the Sigmod function to obtain the attention map. Then we integrate the obtained attention map into the network again to enhance the attention to the texture edge structure of the feature map. We suggested four AMRRs in series for low-light images enhancement (renamed LE-AMRRs). We validated our method on standard low-light test datasets. The experimental results show that LE-AMRRs can generate better low-light enhancement results at a smaller computational cost than other current advanced methods.