Edit Propagation Using Deep Neural Network From A Single Image

With the advent of the information multimedia era,Internet technology and digital media technology have developed rapidly,and images and videos,as important carriers of information dissemination,are closely related to people's lives.The popularity of multimedia devices such as cameras,camcorders,mobile phones,and PCs has exposed more and more video image resources,and the processing requirements for visual media have become more and more intense.As one of the research hotspots in the field of computer vision,color editing processing technology is increasingly applied to various fields such as film and television,interactive entertainment,animation creation,remote sensing communication,and medical imaging.Although more and more researchers are working on color processing technology,there are still key problems in color images such as color unevenness,border color mixing,coloring efficiency,and color overflow.With the development of deep learning,deep neural network(DNN)has been widely used in image semantic classification,target object detection,scene image semantic segmentation and other tasks and achieved breakthrough results,because DNN's supervised end-to-end The layer features self-learning methods and the powerful feature expression capabilities demonstrated.Due to some difficult problems in traditional editing and propagation algorithms,the method of editing and disseminating images using neural networks has become the focus of research.In order to improve the efficiency of image color editing and improve the quality of color editing,this paper proposes a color editing algorithm based on single image training neural network,which is suitable for color editing with various complex natural images.The research work in this paper mainly includes the following points:(1)It is proposed to preprocess the image by using the Euclidean formula,convert the image space feature(coordinate)into a distance map,and cascade with the visual features of the image to generate a multi-channel image,effectively combining the visual features of the image and Spatial features,and obtaining a sufficient number of sub-blocks from the multi-channel image as a training sample set for the DNN network.(2)Construct a new neural network,and use the multi-channel image sub-block as the input of the neural network,automatically extract the depth features that conform to the user interaction and train the deep neural network end-to-end,and image features through the neural network.The weight between the two is automatically assigned,and the traditional method needs to repeatedly perform the manual adjustment of the weight between the image features.(3)A new editing propagation method is proposed.The trained network is used to classify the image,and the image pixels are estimated to belong to the probability value of each type of user interaction,and the classification probability map of the image region is obtained.Post-processing the probability map with the conditional random field can effectively eliminate the influence of the image misclassification pixel,further improve the accuracy of the probability maps and effectively match and adjust the gray channel between the image and the interactive color.Get high quality color editing results.