Method and system for weakly supervised image semantic segmentation based on node classification

Abstract

The invention discloses a weakly supervised image semantic method and system based on node classification, aiming at transforming image semantic segmentation into a graph node classification problem. The method includes: inputting an original image into a classification network, and using image-level labels to obtain initial pseudo-labels; The initial pseudo-label is applied to node classification, and after training and CRF post-processing, the final image segmentation training label is obtained; the image is input into the segmentation network, and the optimized segmentation label is used for training to obtain the final result. The method provided by the present invention converts semantic segmentation into a node classification problem, introduces a graph model, fully considers the relationship between each pixel, and significantly improves the accuracy of a weakly supervised semantic segmentation model when only image-level annotations are used.

Description

technical field [0001] The invention relates to the field of computer vision, in particular to a method and system for semantic segmentation of weakly supervised images based on node classification. Background technique [0002] Semantic segmentation is an essential task in computer vision that aims to identify a class for each pixel in an image. Applied to various scenarios such as autonomous driving and intelligent medical care. Supervised learning based on convolutional neural networks has made significant progress in semantic segmentation. However, it is challenging to obtain fine-grained dataset labels, which are very expensive and time-consuming, and thus difficult to obtain in large batches. In contrast, weakly supervised datasets are easier to obtain. Therefore, it is of great significance to explore effective weakly supervised semantic segmentation methods. [0003] Classification models are widely used in weakly supervised semantic segmentation tasks due to the...

AI Technical Summary

Problems solved by technology

However, classification models can only find the most discriminative part of an object, not the entire object

Therefore, the segmentation accuracy of this method lags far behind fully supervised methods

Most existing models such as FCN and U-Net suffer from the limitations of convolution operations and cannot capture a more comprehensive relationship between pixels, which leads to a series of problems such as incomplete target prediction and inaccurate positioning of semantic segmentation.

Images