Natural image matting method based on deep learning

Abstract

The invention discloses a natural image matting method based on deep learning, and the method comprises the following steps: obtaining a matting data set, and carrying out the data enhancement; establishing a natural image matting model with an encoder-decoder structure; in order to reserve detail information, designing an encoder to enable the downsampling multiple to be 4, in order to compensatefor reduction of a receptive field caused by downsampling multiple reduction, introducing a cavity convolution to expand the receptive field, and storing the maximum pixel position in the maximum pooling operation so as to provide position information for an upsampling stage; in order to solve the multi-scale problem, a cavity space pyramid module is connected to the top of the encoder; designinga global context module in a decoder, wherein the global context module is used for fusing high-level features corresponding to the encoder and the decoder; finally training and testing. According tothe method, more detail information is reserved in the feature extraction process, and meanwhile, multi-scale features are associated, so that the model can capture global information, model processing details and large-area transparent objects are facilitated, and the matting quality is improved.

Description

technical field [0001] The invention relates to the technical field of image processing, in particular to a natural image matting method based on deep learning. Background technique [0002] Cutout technology is an image processing technology on digital images. It was originally developed by the film and television industry, and has now become a crucial technology in the production of visual effects. Using matting technology, producers in the fields of movies, advertisements, posters, etc. can seamlessly embed a desired character or object into a specified scene. However, the production of such special effects mostly uses the blue screen cutout technology, which needs to place the characters or objects to be embedded in a solid color background for shooting, which greatly limits the application of the cutout technology. With the development of computer technology, users have an increasingly strong demand for extracting objects of interest from a natural image, and at the sa...

AI Technical Summary

Problems solved by technology

However, if the downsampling factor is simply reduced, the receptive field will be too small, and the feature extractor can only learn local features in a small range, which is not conducive to the acquisition of high-level semantic information and also causes a decrease in accuracy. Contradictions between extracting high-level semantic information

For problems such as missing blocks and low prediction accuracy of difficult samples, the targets of map matting are generally relatively large. These large targets can be regarded as composed of blocks of different sizes. If only convolution operations with the same kernel size are used, the same experience In the wild situation, it is easy to cause different features of large blocks to be extracted due to different internal details of the blocks, lacking higher-level overall features, and several adjacent small blocks may be recognized as a whole, so the image matting The multi-scale problem of objects also needs to be solved in the task; in addition, in order to solve the problem of missing blocks in large-area transparent foreground and the problem of difficult samples, the natural image matting model also needs to introduce a method to expand the scope of context information

Images