Reality scene image segmentation method based on contrast self-supervised learning

Abstract

The invention relates to a real scene image segmentation method based on contrast self-supervised learning. The method comprises the following steps: designing a contrast self-supervised learning segmentation model and a loss function; the comparison self-supervised learning segmentation model comprises an upper branch and a lower branch, and each branch comprises an encoder, a decoder, a feature projection module and a predictor which are connected in sequence; an input image is randomly cut into two image patches with an overlapping area, and the two image patches are processed by comparing two branches of the self-supervised learning segmentation model; the designed loss function comprises image-level loss and pixel-level context alignment loss, and the image-level loss is mainly used for calculating loss between an output feature map of the feature projection module and a prediction map obtained by the predictor; pixel-level context alignment loss mainly constructs a spatial context relationship between image pixels by maximizing the similarity of two image patch overlapping regions and minimizing the interference degree of a non-overlapping region on feature extraction.

Description

technical field [0001] The invention relates to the technical field of machine vision, and specifically provides a method for segmenting real scene images based on contrastive self-supervised learning. Background technique [0002] With the intelligent development of automobile technology, unmanned driving has become a research hotspot. Unmanned vehicles usually need to be equipped with various sensors to analyze the driving environment by collecting images of real scenes during driving, and then guide unmanned vehicles to drive safely. In the process of unmanned driving system design, image analysis and scene understanding are very important links. The key to image analysis is the segmentation of real scene images. [0003] Although image segmentation based on deep learning has achieved good results, it needs to rely on a large amount of manually labeled data, which is costly and time-consuming. Self-supervised learning is a representative branch of deep unsupervised lear...

AI Technical Summary

Problems solved by technology

However, in image segmentation, the model does not know how many meaningful regions should be generated in the image, that is, the loss of spatial continuity is not considered. For details, see the literature "Roh B, Shin W, Kim I, et al.Spatially consistent representation learning [C ] / / Proceedings of the IEEE / CVF Conference on Computer Vision and Pattern Recognition.2021:1144-1153.》

[0004] In addition, existing self-supervised learning techniques usually focus on generating image-level global invariant representations, ignoring the relationship between pixel contexts within the image, thus limiting the generalization ability of the model for image segmentation, resulting in unsatisfactory segmentation results

Images