Attention-Separable Convolutional Network Model For Lowlight Image Enhancement

Images captured in low illumination environments often suffer from low contrast,blurred images and loss of detail,which can affect the usability and readability of the image and in turn can have an impact on higher level computer vision tasks.Significant progress has been made in deep learning-based models for low-light image enhancement,but the current trend is to create deeper,more complex networks to improve accuracy.However,in many practical applications,improving accuracy does not necessarily make the network more efficient in terms of size and speed.Furthermore,for some extreme cases,such as poor sharpness,loss of detail and overexposure,current methods are still not good enough to improve the overall quality of the image.Therefore,two solutions are proposed in this paper to address these problems as follows:(1)This dissertation presents a low-light image enhancement algorithm based on an attention mechanism and depth-separable convolution,aiming to improve the sharpness and contrast of images.Firstly,a depth-separable convolutional enhancement factor extraction network is designed to effectively estimate the pixel-level light deficiency of low-light images.Secondly,a recursive image enhancement network combined with an attention mechanism(SE-RIE)is designed to progressively enhance low-light images with moderate model size.Finally,the enhanced images from SE-RIE are feature extracted and the high semantic layer is converted into a high resolution layer for semantic segmentation.Experiments demonstrate that the algorithm proposed in this paper can significantly improve image quality,avoid colour distortion and obtain better subjective and objective evaluation metrics.The use of deep separable convolution reduces the model complexity,while the attention mechanism helps the network to better focus on the important parts of the image and improve the enhancement effect.(2)This dissertation proposes a new image enhancement model CUCEAT-Net,which aims to address the problem of missing detail information and edge blurring phenomenon in low-illumination images.Based on the U-Net architecture and the principle of iterative enhancement,the model uses a feature fusion module to effectively fuse global and local information to solve the problem of unclear generated images.To further improve the performance,this paper introduces a colour enhancement module that fuses attention mechanisms to improve the expressiveness of features and adaptively enhance images in different luminance regions.In experiments,the algorithm proposed in this paper can significantly improve the sharpness of low-illumination images,solve problems such as blurred edges and recovered details,and achieve better enhancement results.