Semi-supervised image semantic segmentation method and device based on self-supervised low-rank network

Abstract

The invention discloses a semi-supervised image semantic segmentation method and device based on a self-supervised low-rank network, and the method comprises the steps that: the self-supervised low-rank network is constructed, the inverse geometric transformations are performed masks from two branches, a pseudo mask is generated through an optimization module, and the pseudo mask is input into an LR low-rank module; in each iteration, an assignment matrix P is calculated through softmax normalization attention and a temperature coefficient; the optimal basis mu is updated by aggregating the input feature X, and after a softmax normalized class activation graph A with the class being C and a deep feature X1 are obtained, the kth initialization basis is calculated through a weighted average value; and, in the base initialization process, a target function composed of classification loss and pseudo mask segmentation loss is used for supervision, an output result of an LR low-rank module is decoded and optimized, and the self-supervision low-rank network is updated according to the loss. The device comprises a construction module, an optimization module, an LR low-rank module, an updating module, a prediction module, a supervision module and an output module.

Description

technical field [0001] The invention relates to the field of image semantic segmentation, in particular to a semi-supervised image semantic segmentation method and device based on a self-supervised low-rank network. Background technique [0002] Recently, deep learning-based semantic segmentation models have achieved significant progress by training with large-scale pixel-level labels. However, this supervised approach requires extensive human annotation, which is time-consuming and expensive. In order to reduce the workload of annotating pixel-level labels, a large number of studies have developed Weakly Supervised Semantic Segmentation methods (WSSS) with low-cost annotations, such as: bounding boxes, scribbles, points, and image-level labels. [0003] Most of the popular image-level WSSS methods need to go through multiple training and optimization stages to obtain more accurate pseudo-labels. These methods usually start with weakly supervised localization, e.g. Class A...

AI Technical Summary

Problems solved by technology

However, this also increases the risk of error being imitated and accumulated, and there will be a phenomenon of gradient flow backpropagating from the top layer to the lower layer

Images