Image information extraction and generation method based on inter-region attention mechanism

Abstract

The invention provides an image information extraction and generation method based on an inter-region attention mechanism. The method comprises the following steps: 1, a generator coding stage: takinga semantic label graph in a training set as an input, dividing the semantic label graph into four branches for coding, and obtaining four branch images; step 2, a generator decoding stage: splicing the four branch images generated in the encoding stage to generate a false image as input of discriminator training; 3, a discriminator training stage, splicing the semantic label graph in the trainingset with the real image of the training set and the false image generated by the generator on the channel dimension to serve as input of the discriminator; dividing the discriminator into two scaleswhich are respectively an original scale and a downsampled two-time scale; on each scale, inputting a downsampling convolution layer which continuously passes through four layers and finally passes through one layer of convolution, and outputting 0 or 1 at each position to serve as true or false of a prediction image, wherein namely 0 represents that the current position is predicted to be a falseimage, and 1 represents that the current position is predicted to be a true image.

Description

technical field [0001] The invention is an image information extraction and generation method based on an inter-regional attention mechanism, which belongs to the field of computer vision image information extraction and image generation. Background technique [0002] With the in-depth application of big data technology and the continuous improvement of CPU and GPU computing capabilities, deep learning has received extensive attention in computer vision, data processing, natural language applications, and autonomous driving. [0003] However, since the birth of deep learning, the lack of data sets has been hindering the development of deep learning. In order to solve the problem of lack of data sets in the image field, image generation has attracted the attention of the industry and has developed rapidly. Not only that, image generation also has very broad application scenarios and research significance in many other important fields, such as navigation, image color transfo...

AI Technical Summary

Problems solved by technology

First, long-distance dependencies are difficult to obtain

That is to say, although the receptive field of convolution is constantly expanding after continuous subsampling (generally 3 layers), the high-definition image has high precision, generally reaching 1024×512, so it is difficult to capture the global corresponding dependence relation

Moreover, after continuous underground sampling, the features of different places in the image are too mixed, making it difficult to generate fine images

Secondly, for a 1024×512 image, different places have corresponding semantic relations, but because the features are too mixed, it is difficult to achieve accurate generation of the same semantic relations in different places

In addition, in the overall generation, the generation of a certain pixel will inevitably be affected by other positions, and in the global scope, it will inevitably lead to a uniform result, that is, pixel blur

Images