| In dark and cloudy environments,underexposed low-light images are easily obtained.The low overall or local brightness and lack of detail in these images seriously affects the image quality and has a detrimental effect on both the visual perception and the subsequent visual task.The aim of low-light enhancement techniques is to enhance the overall brightness of the image and restore the details in the dark areas,thus restoring the low-light image to a normally exposed image.To deal with low-light images under complex illumination,this paper conducts research based on convolutional neural networks and investigates the lowlight image enhancement task in two aspects: the enhancement model and the network structure.The main research contents are as follows:(1)Proposed algorithm for low-light image enhancement based on image-to-patch model.To solve the problem of enhancing low-light images at different luminance levels,an image-to-patch enhancement model is proposed,which indirectly enhances the image by introducing a patch image and guiding the convolutional neural network to fit this intermediate image.The patch image preserves the incremental information from the lowlight image to normal,allowing the network to focus on the dark areas of the image and reduce the attention to the normal areas.In addition,a lightweight network structure is designed to minimize the time and resources required for enhancement while maintaining good enhancement results.Finally,the effectiveness of the method is demonstrated through experimental analysis.(2)Retinex low-light image enhancement method with fused color equalization is proposed.To solve the problem that the traditional Retinex theory-based method can easily lead to color distortion and loss of detail information,it is proposed to restrict the Retinex theory to the greyscale domain of the image.The incident and reflected components are obtained separately in the grey-scale channel by designing a decomposition network.The reflectance is fed into the reconstruction network along with the input image as a complement to the detail and structure.The reconstruction network is designed with a skip connection and color balance module to enhance the reconstruction quality.The color equalization module consists of a channel attention mechanism and point-by-point convolution to maintain the color stability of the output image.In addition,a segmented loss function is used in the training phase to improve the performance of the network The effectiveness of the relevant modules is demonstrated through relevant ablation experiments.By comparing with other methods,the proposed method has satisfactory results in terms of objective metrics and subjective perception. |
