A Multi-style Generative Adversarial Network for Underwater Image Synthesis and Its Application

Abstract

The present invention provides a multi-style generation confrontation network underwater image synthesis method and its application, collecting RGB-D images on land, constructing a land RGB-D image data set, and collecting underwater images of different styles as a real multi-style For underwater domain images, build a generation confrontation network model based on CycleGAN, input land domain images and underwater domain images into the network model, and convert land domain images into synthetic multi-style underwater domain images through training and iterative feedback. The synthesized multi-style underwater domain images contain features such as texture and color of underwater real images. In addition, the present invention inputs the synthesized multi-style underwater domain image and the land depth image in the RGB-D image data set as an underwater RGB-D training data set into a supervised depth estimation network to obtain an underwater scene depth estimate, It provides a basis for further research on the water environment.

Description

technical field [0001] The invention relates to the technical field of underwater computer vision, in particular to an underwater image synthesis method of a multi-style generation confrontation network and its application. Background technique [0002] Underwater vision is the basis for the study of oceans, lakes and other waters. In a complex water environment, it is difficult to obtain image datasets with water information with equipment. Due to the strong absorption and scattering effects, underwater imaging and analysis have certain limitations, which restricts the development of underwater visual depth estimation and other technologies. [0003] Synthesizing an underwater dataset of specific water quality parameters from a land dataset matched with depth information plays a vital role in underwater vision research. However, there are currently few underwater image synthesis methods, and the conditions for synthesizing underwater images require corresponding water qua...

AI Technical Summary

Problems solved by technology

[0004] The present invention provides a multi-style generative confrontation network underwater image synthesis method and its application to solve the problem that the existing underwater image synthesis methods are few, and the conditions for synthesizing images are relatively high, and multi-style underwater images cannot be synthesized. Technical issues, the synthesis method converts land images into underwater images of different styles, and estimates the depth of underwater scenes, which provides a basis for further research on the water environment, such as based on the depth estimation of underwater scenes, And then provide a basis for the research of underwater robot path navigation and 3D reconstruction of underwater scenes

Images