Research On Gradient Domain Filtering Method Based On Unsupervised Learning

Edge-preserving filtering is an important technique in the field of modern computer photography and an important topic in the field of computer graphics.The initial image often contains a lot of distracting information that can degrade the quality of the image and interfere with the acquisition of information.The aim of edge-preserving filtering is to smooth out low-contrast details and remove distracting information from the image,thereby obtaining the main information and structure of the image.Existing edge-preserving filters can be classified as local edge-preserving filters,global edge-preserving filters and deep learning approximate edge-preserving filters.Local edge-preserving filters are highly efficient but have serious problems with artefact halos;global edge-preserving filters have excellent image quality but are inefficient,computationally expensive and not suitable for real-time information processing;deep learning approximate edge-preserving filters are efficient but not flexible enough to use and also have problems with image artefact halos.There is still much room for research on the effectiveness and computational efficiency of edge-preserving filtering.Based on an in-depth study of the above problems,the text proposes a study of gradient domain filtering methods based on unsupervised learning.The specific work is as follows.（1）In order to improve the efficiency and image quality of edge-preserving filtering,this paper chooses an unsupervised approach to the study of edge-preserving filtering,which is combined with a weighted least squares framework to propose an edge-preserving filter under a weighted least squares framework based on deep unsupervised learning.Among the objective functions,this paper uses weights to protect the apparent gradients in the image to enhance the edge-preserving capability.And based on the spatially varying smoothing property of the edge-preserving filter,this paper proposes a lightweight fully convolutional neural network based on dilated convolution with different expansion factors.The proposed filter makes full use of 2D neighborhood information and is therefore able to suppress various artefacts.Due to the highly optimized framework of deep learning,the proposed filter is very efficient,capable of processing 720P images at 0.085s on 1080.In order to overcome the fact that filters in deep learning usually have only one effect and cannot be changed flexibly,this paper further proposes interpolation-based multi-scale image prediction on unsupervised filtering to make our filter more flexible in order to get more results at different scales to accomplish more image processing tasks.（2）To improve the edge-preserving capability of the edge-preserving filter and further enhance the execution efficiency,a weighted and truncated L₁filter based on unsupervised learning is proposed by modelling the filtering task as an L₁normalization model.The weighting and truncation functions in this method significantly improve the edge-preserving properties of the filter.A deep unsupervised learning-based approach is also proposed to solve this optimization model.The proposed solution exploits the U-Net structure and uses it for multiscale image processing,making the filter more flexible.The results show that the filter outperforms state-of-the-art filters in terms of image quality on a variety of tasks,such as image smoothing,detail enhancement,HDR tone mapping and edge detection.At the same time,the filter is very fast.Capable of processing 720p images at 0.061s on 1080.Both qualitative and quantitative experiments demonstrate the advancement of the described method.（3）This paper designs and implements a prototype system for image processing based on unsupervised learning of gradient domain filtering methods.The system has information management,user interaction and algorithm modules,and mainly implements a variety of image processing tasks applicable to the two methods mentioned in this paper.The interface is simple and easy to use,and the user can adjust the parameters to obtain different image results,which verifies the theoretical significance and practical value of this research.