Self-supervision-based visual representation learning method

Abstract

The invention provides a self-supervision-based visual representation learning method. Main contents of the method include: self-supervision, visual representation, image construction and transformation learning in an image. A process of the method includes: adopting a self-supervision method to learn invariance of which expression includes differences between instances and differences inside theinstances; building a graph, in which nodes indicate image patches, describing similarities between the image patches, and defining two types of edges associating the image patches mutually in the graph, wherein constructing the graph with the instances and the internal edges includes scaling moving objects and is through the edges between the clustered instances and the edges inside the tracked instances; and training deep neural networks to generate similar visual representation. The visual feature learning method of the invention carries out self-supervision, can automatically acquire annotation labels, and greatly saves manpower and material resources; and reduces, at the same time, impacts brought by viewpoints, postures, deformation, illumination and the like, and increases an accuracy rate.

Description

technical field [0001] The invention relates to the field of visual representation learning, in particular to a self-supervision-based visual representation learning method. Background technique [0002] As the society enters the digital information age, the rapidly increasing amount of images and videos brings great challenges to data management and analysis, which makes intelligent visual data classification and retrieval technology attract more and more attention. Visual representation learning is to use cameras and computers instead of human eyes to recognize, track, and measure targets, and then do further graphics processing, and use computers to process images that are more suitable for human eyes to observe or send to instruments for detection. It can be applied to visual object recognition, such as automatic labeling of Web images, massive image search, image content filtering, medical remote consultation and other fields; it can also be applied to the detection of ...

AI Technical Summary

Problems solved by technology

[0004] Aiming at the problem that it is easily affected by viewpoint, pose, deformation, lighting, etc., and the accuracy is not high, the purpose of the present invention is to provide a self-supervised visual representation learning method, which adopts the self-supervised method to learn invariance, and its representation includes examples The difference between the differences and the difference within the instance, construct a graph describing the similarity between the image patches, the nodes in the graph represent the image patches, define two edge types that relate the image patches to each other in the graph, and construct a graph with instances and internal Graphs of edges, including scaling moving objects, edges between instances via clusters, and edges within instances via traces, train deep neural networks to generate similar visual representations

Images