Abstract:
Abstract： Low Light Image Enhancement （LLIE）， which is to restore images captured under insufficient lighting conditions to normal exposure images. The existing LLIE algorithms based on deep learning often use stacked convolution or up/down sampling methods， which lacks the guidance of relevant semantic information， resulting in problems such as increased noise， color distortion and detail loss in the enhanced image. To address this issue， a novel LLIE algorithm based on dual attention residual modules is proposed. This algorithm proposes a residual block that integrates dual attention units（Dual Attention Residual Block， DA-ResBlock）， which provides semantic information guidance in both channel and spatial domains. Through multi-level cascaded DA-ResBlocks， effective features are stably extracted， and skip connections and convolutional neural networks are used to restore image detail information. In addition， a composite loss function is used to constrain the enhancement task. Finally， we compare our algorithm with mainstream algorithms in recent years on two public datasets that provide real images. The experimental results show that the proposed algorithm effectively improves image brightness while better suppressing noise， restoring image color and detail texture in subjective vision. In the objective evaluation， the three indexes of PSNR， SSIM and LPIPS are superior to the compared mainstream algorithms.

Key words: Key words： image enhancement, low-light image, visual attention, residual network