Weak supervision image semantic segmentation method, system and device based on cross-image association

Abstract

The invention belongs to the field of deep learning, computer vision and pattern recognition, particularly relates to a weak supervision image semantic segmentation method, system and device based oncross-image association, and aims to solve the problems of incomplete semantic segmentation targets, inaccurate positioning and inaccurate semantic category judgment caused by rough annotation adoptedby weak supervision. The method comprises the steps: acquiring complementary information from images of multiple objects of the same category through a cross-image association relation module, and obtaining fusion features; training an image semantic segmentation model based on the fusion features; and through the trained model, obtaining a semantic segmentation result of a single input image ora plurality of same-class object image groups. According to the invention, the pixel-level pseudo image annotation is generated from the rough weak image annotation, and the complementary informationis obtained from different images in the model training process, so that the defect of incomplete pseudo image annotation is overcome, and the performance of the weak supervision semantic segmentationmodel can be remarkably improved under the condition of only depending on the image-level annotation.

Description

technical field [0001] The invention belongs to the fields of deep learning, computer vision and pattern recognition, and in particular relates to a method, system and device for semantic segmentation of weakly supervised images based on cross-image association. Background technique [0002] Semantic segmentation is a basic task in computer vision. Its goal is to label each pixel in a picture with a corresponding semantic category. It plays an important role in subsequent image understanding, image editing and other tasks, such as pedestrian segmentation in traffic scenes. , medical scene lesion segmentation, military scene aerial image segmentation, etc. At present, more mature semantic segmentation technologies are basically based on deep learning methods. However, the training of deep learning networks usually requires a lot of manpower and time to complete the fine manual annotation of training images, which is very time-consuming and labor-intensive. It is difficult to...

AI Technical Summary

Problems solved by technology

[0005] In order to solve the above-mentioned problems in the prior art, that is, the problems of incomplete semantic segmentation, inaccurate positioning, and inaccurate judgment of semantic categories brought about by rough annotation adopted by weak supervision, the present invention provides a weak Supervised image semantic segmentation method, the image semantic segmentation method includes:

Images