A cross-domain large-scale scene generation method

Abstract

The invention discloses a cross-domain large-scale scene generation method, which belongs to the technical field of image generation. First, the corresponding distribution of each background in the target scene and the semantic features of each foreground target are set; The background segmentation map, and then each foreground object is added to the background segmentation map in turn according to the semantic features; then, the segmented image is fused with a pre-set feature vector; the image after feature fusion, using the multi-volume neural network Concatenate layers for coding, extract high-level features, obtain high-level semantic feature maps, and use residual network to further fuse them; finally, use the upsampling structure to decode the fused residual results; finally output the color scene generation results. The invention provides a data generation method for a large-scale viewing angle task where data is insufficient and samples are difficult to obtain, and the process of manually giving a segmentation and labeling map is omitted, and the effect is good and the practicability is stronger.

Description

technical field [0001] The invention belongs to the technical field of image generation, and relates to a cross-domain large-scale scene generation method. Background technique [0002] In recent years, as the country continues to open up the right to use low-altitude areas, aircraft are often used to perform tasks such as auxiliary disaster inspections and special scene surveillance. In these tasks, the acquisition and analysis of scene information is one of the most important links. However, it is the first contact with many scenes in actual operation, and lack of prior experience brings great challenges to the analysis of information. The scene generation method can simulate the characteristics of the target scene, and has an immeasurable effect on the situation where the scene data is difficult to obtain in advance or the data is missing. [0003] At present, many deep learning tasks require a large number of samples as training sets, and the algorithms used are also to...

AI Technical Summary

Problems solved by technology

However, this method requires a segmented annotation map of the same size as the output because the entire network input requires pixels of the same value to represent the area where each type of object is located. limit its application

Images