Weak supervision semantic segmentation method and device based on context decoupling and data enhancement

Abstract

The invention discloses a weak supervision semantic segmentation method and device based on context decoupling and data enhancement. The method comprises the following steps: inputting an image data set into a weak supervision semantic segmentation model to obtain a semantic segmentation mask; according to a set standard, selecting object instances from the semantic segmentation masks to form a mask set; randomly pasting foreground objects in the mask set into the image data set by adopting an online enhancement mode to obtain an enhanced image set; and jointly inputting the image data set and the enhanced image set into a weak supervision semantic segmentation model to obtain a final semantic segmentation result. According to the method, context association information between a foreground and a background in weak supervision semantic segmentation is considered, and the association information can be decoupled from the image, so that the network is more concentrated on the foreground, additional data is not needed, and the segmentation performance is improved; a good solution is provided for scenes such as medical data analysis and automatic driving in practical application.

Description

technical field [0001] The invention belongs to the technical field of semantic segmentation and data enhancement, and in particular relates to a weakly supervised semantic segmentation method and device based on context decoupling and data enhancement. Background technique [0002] In recent years, with breakthroughs in computer hardware and deep learning, the field of artificial intelligence has developed rapidly in more and more fields. As the most extensive field of artificial intelligence research, computer vision includes image classification, object detection, image segmentation, image generation, OCR text recognition and many other detailed directions, and has broad applications in face recognition, driverless driving, image retrieval and other fields . [0003] Image segmentation is a major branch of computer vision, dedicated to judging the category of each pixel in the picture, but at the same time, because the fully supervised image segmentation mask requires a ...

AI Technical Summary

Problems solved by technology

[0004] Although object area expansion techniques emerge in endlessly, they all use CAM as the cornerstone; the effect of subsequent expansion is based on the segmentation results of CAM in the first step; however, the supervision information accepted by CAM is only image-level labels, when the object is related to the context When the degree is high, such as boats and water, planes and sky, trains and railroad tracks, CAM will mistakenly identify the background as part of the foreground, so that the trained network does not focus on the foreground, but other related strong contextual information as the basis for discrimination

Images