| Image contrast enhancement is an important research topic in the field of computer vision,and its demand originates from practical problems in real life.Contrast enhancement of images can improve their quality and information content,which helps them be better applied.However,existing image contrast enhancement algorithms have certain limitations.This thesis conducts research on the issue of image contrast enhancement.The main work of the thesis is as follows:(1)An image contrast enhancement algorithm based on attention mechanism and Retinex model is proposed.The local and global features are extracted respectively,and the convolutional neural network is trained to predict the local affine transformation coefficients in the bilateral space,so that the network has the ability of local adaptive enhancement.Extract the reflection components of the image on a full resolution image to better guide subsequent interpolation work.(2)In order to solve the problem of image overexposure and color deviation,an exposure protected image contrast enhancement algorithm is proposed.This algorithm performs gamma transformations on input images with different parameters to obtain images under various exposure conditions,thereby enhancing the generalization ability of the network;Statistics and analysis of mapping relationships on the dataset,and use statistical results to constrain the output of the network;Transforming the color space and separating channels,processing them separately in different channels,solves the problems of brightness enhancement and color deviation.(3)Build a paired dataset of natural images.By organizing and supplementing the existing image enhancement dataset,a set of input images were obtained.After software processing and manual adjustment,the input images were used as the target images,and a natural image paired dataset containing multiple scenes was constructed for model training and testing.The algorithm proposed in this thesis achieved a peak signal-to-noise ratio of 26.43 and a structural similarity of 0.90 on the test set,which has been improved compared to current mainstream algorithms. |
