Color- And Texture--driven Image Content Recreation

As an intuitive and easy-to-understand information carrier,images are widely used in digital media,computational agriculture,security and detection,and art design.In general,these applications require processing image content,and the human visual system is sensitive to both color and texture,so the topic of color-and texture-driven image content recreation has become a key research direction in the field of computer graphics and image processing.This dissertation presents an extensive and in-depth study of the key techniques and theories involved in the application of color and texture as clues.The research topic of how to perform reliable content recreation of images faces many challenges.For example,in the application of image color content recreation,how to ensure that the visual effect of color editing conforms to human visual perception,how to improve the convenience of the interaction process and how to efficiently evaluate different color editing methods are all relatively challenging issues.For example,in the application of image texture content recreation,how to use the characteristics of the image itself to design a targeted convolutional neural network model and how to ensure that the model has a high degree of generality are also key issues to be solved.Given that the human visual system is highly sensitive to color and texture,this dissertation studies the image content recreation problem from several perspectives with color and texture as carriers,and proposes a series of image color editing,texture restoration and texture generation algorithms.In this dissertation,we first propose a novel image recoloring method by taking the difficulty of color editing in shadow regions of images as the starting point.Then,the representation of highly dynamic images is explored,and the representation space suitable for color transfer of highly dynamic images is constructed by combining representation learning with self-supervised learning in a pioneering way.Finally,for the research topic of image texture content recreation,the two problems of texture restoration and generation in images are investigated,and it is found that the targeted design of convolutional neural networks combined with the nature of the target problem can enhance the model effect.Specifically,the research in this dissertation is as follows.(1)proposed an translucent image recoloring method based on the homography transformation.The method proposes a clustering method based on histogram statistics for extracting the dominant colors that can represent all the colors of an image.Then the homography transform is used to spread the user's modification of the dominant color to all color changes of the image,and finally the problem that the color of some pixels is out of the color gamut is solved by using a nonlinear optimization technique.In addition,in this study,a method for synthesizing ground truth results based on ray-tracing techniques is proposed to quantitatively evaluate different image recoloring methods.(2)proposed a highly dynamic image color transfer method based on representation learning.The method combines classical color transfer algorithms to design novel adversarial neural networks and designs self-identity loss and adversarial loss to train this network based on the idea of self-supervised learning.Under the same color transfer method,the representation learned by the neural network proposed in this dissertation can obtain more natural and beautiful color transfer effect.(3)proposed a multi-scale transformation based PNG-8 image texture restoration method.The method addresses the problem of uneven distribution of dithering noise in images and designs global and local transformation network modules based on the idea of multi-scale transformation.The global transformation module can learn a set of explicit optimal convolution kernels from the training dataset,while the local transformation module can generate implicit specific convolution kernels adaptively according to the differences in the distribution of dithering noise in samples,and the neural network composed of these two modules outperforms in both quantitative and qualitative evaluation experiments outperform the existing methods.(4)proposed a method for smoothed image texture generation based on visual perceptual properties.The method is based on the spatial variability and spatial correlation of images,and creatively designs the spatial feature transformation module and the full attention module.Among them,the spatial feature transformation module can generate different texture information for different regions,while the full attention module can consider the association between each region to make up for the respective textures.In comparison with existing methods,this dissertation achieves the best results in both quantitative and qualitative evaluation experiments.This dissertation presents an study of color and texture-related applications from multiple perspectives,focusing on the direction of image content recreation,and proposes detailed solutions to the difficulties involved.Through quantitative and qualitative evaluation experiments,it is found that the proposed algorithms outperform existing methods in their respective subfields and can be extended to relevant application scenarios.