Deep learning matting method based on synthetic data set augmentation

Abstract

The invention discloses a deep learning matting method based on synthetic data set augmentation, and the method comprises the following steps: carrying out the data augmentation of an adobe data set through DAZ3D software, and synthesizing a data set needed by deep learning; performing morphological operation of corrosion and expansion on the alpha mask in the data set to obtain a tripartite graph required by training; constructing a network structure suitable for matting on the basis of a VGG16 network structure, carrying out convolution on four-channel input formed by splicing an image and a tripartite image by utilizing a coding and decoding structure of the VGG16 network, and outputting a rough matting result; constructing a network structure for fine cutout, splicing the obtained rough cutout result and the source image, carrying out convolution to obtain a prediction result with a clear boundary, combining the rough cutout training to form an integral network, repeatedly training the integral network, and updating the weight of the integral network; and storing the obtained weight of the whole network as a pre-training network model for the requirement of subsequent batch matting. According to the invention, the matting of the hair-level precision of the image with the natural background is realized.

Description

technical field [0001] The invention relates to the field of computer image processing, in particular to a deep learning matting method based on augmentation of synthetic data sets. Background technique [0002] Image matting is a field with a wide range of application backgrounds. From image editing software to film and television special effects production, the problem of background matting is involved. Image matting and image segmentation belong to a large class of problems. The goal of image segmentation is "pixel-level" accuracy, while the goal of matting is "half-pixel"-level accuracy, that is, it is necessary to solve the corresponding transparency of foreground objects in the image. Both have Very high similarity. Although the problem of image segmentation has been studied more maturely, if there are details of soft edges such as hair on the edge of the foreground, or the foreground object has characteristics such as translucency and refraction, using the cutout ope...

AI Technical Summary

Problems solved by technology

The invention uses semantic segmentation information to guide the design of the matting method, but the fineness of the model does not reach the hair-level precision matting

Images