Damffn: Depth-Wise Convolution Attention and Multi-Scale Feature Fusion Network for Low-Light Image Enhancement

Abstract

Images obtained by existing low-light image enhancement methods still suffer from poor visibility, low contrast, and loss of spatial details. This paper proposes a depth-wise convolution attention and multi-scale feature fusion network for low-light image enhancement (DAMFFN) to solve these problems. In DAMFFN, firstly, this paper designs a low-light attention block (LLAB) consisting of a low-light multi-head self-attention block (LL-MSAB), a dual-branch equalization block (DBEB), and two layers of normalization composition. The LL-MSAB balances feature weights between different channels by calculating the attention map between channels, thus improving the image's visibility and contrast. Secondly, this paper leverages the DBEB to enhance the image's contrast. Finally, the paper proposes a multi-scale feature compensation block (MSFCB), which is used to reduce the loss of spatial details in the LLAB and downsampling stages. On the other hand, MSFCB can also fuse the deep spatial information of images of different scales. In terms of the loss function, this paper refers to a multi-scale frequency domain loss function (Mult-SFD), which reduces the difference in the frequency domain space between the reference image and the enhanced low-light image. This paper conducts sufficient qualitative and quantitative experiments in 5 public datasets, which are superior to many other low-light image enhancement methods in terms of visual effects and index scores(PSNR: Peak Signal-to-Noise Ratio. SSIM: Structural Similarity Index): PSNR=24.87dB, SSIM=0.856 on LOw-Light(LOL) dataset; PSNR=25.70dB, SSIM=0.912 on MIT-Adobe FiveK dataset.

Keywords: Low-light image enhancement, Low-light multi-head self-attention, Multi-scale Feature compensation, Dual-branch equalization, Frequency domain