Research On Unsupervised Image Enhancement Method Based On Attention Mechanism

Image enhancement is a common method in the field of computer vision,which can effectively improve the contrast and brightness of images and thus is of great significance for subsequent visual tasks.However,most existing methods suffer from problems such as excessive brightness,noise enhancement,color distortion,and detail blurring.In this thesis,we systematically study image enhancement methods and introduce the technical principles of related algorithms.We propose corresponding improvement methods for the shortcomings of existing algorithms.The main contents are as follows:In the section on image enhancement techniques,we first propose a gradient field transformation image enhancement method based on an adaptive gamma correction function.This method divides the corresponding gradient intervals according to a threshold and uses an improved gamma correction function for adaptive gradient enhancement in different gradient intervals.In the image reconstruction stage,the algorithm complexity is reduced by using a matrix partition reconstruction strategy.This thesis proposes an unsupervised image enhancement method based on attention map guidance,with LSGAN as the network backbone.The generator uses UNet as the main network and embeds an attention map network and multi-branch convolutional multi-head self-attention module,which are used to guide generator training and enhance the detailed features of images.The discriminator network integrates a convolutional encoder,Transformer encoder,and fully connected neural network to stabilize the entire training process.To objectively evaluate the performance of our method in different scenarios,multiple experiments are set up to evaluate the enhancement results,and the comparison objects involve 14 image enhancement methods.The visualized results show that our method can effectively avoid the relevant problems of existing methods.In addition,our method has outstanding performance in image quality evaluation metrics.Taking the low-light image enhancement comparison experiment as an example,our unsupervised image enhancement method compared with the improved gradient field method and the same type of deep learning method,respectively improved 35.8% and3.72% on the information entropy index.It also ranked first and second in the LOE and NIQE indexes.The application experiment results show that our method has a significant improvement effect on the image-matching task results in low-light and uneven illumination scenarios.