Research On Color Low-light Image Enhancement Algorithm Based On Multi-task Processin

Night vision is a technology that detects nighttime targets through optoelectronic imaging devices.Traditional night vision technologies are monochrome images,and since the human eye can distinguish hundreds of times more colour levels than grey levels,researchers have worked to bring monochrome night vision images to colour.Previously,images produced using signals from different bands were fused into colour images or passed through colour,producing colour images that were unnatural and did not reflect the true colour information presented by the scene during the day.The images obtained by true colour night vision technology are consistent with the colours of objects seen during the day,closer to the true colours of objects and consistent with the subjective vision of the human eye,which can improve the recognition rate of targets and has important research significance and application value.However,truecolour night imaging faces problems such as severe energy loss,significant noise and colour distortion in the reconstructed images.Traditional algorithms usually target a single problem,while deep learning methods can handle multiple tasks together and improve efficiency.A conventional low-light image enhancement method based on image decomposition is proposed for the problems of low brightness,low colour saturation and noise generated by existing true colour night vision cameras.The TV model decomposes the low-light image into a structure image and a texture image;for the structure layer,the light map is extracted by obtaining its luminance component,and then the light map is gamma corrected to obtain a structure-enhanced map according to the Retinex model,which improves the brightness and contrast of the image;then the colour saturation is improved by an adaptive function;for the texture layer,the texture component is also enhanced with the structure component.Finally,the enhanced structure image and the optimised texture image are summed and denoised with BM3 D to obtain the final enhanced result.Numerous experimental results confirm the effectiveness of this method,which can effectively improve image brightness,colour saturation and reduce image noise while preserving details.Deep learning has developed rapidly in recent years and has shown significant advantages over traditional algorithms.Therefore,this paper proposes a neural network algorithm based on channel-calibrated convolution.The upper branch of the channel-calibrated convolution introduces a channel attention block to analyse the features between RGB channels,which is used to replace the traditional convolution in the U-Net network to achieve colour recovery and retain more image information;the Sobel loss function and colour loss function are added to the traditional loss function to suppress noise while preserving image details,reducing chromatic aberration and enhancing contrast.Image data sets from real scenes are collected to improve the processing effect on the actual data.