Small sample training method based on target segmentation

Abstract

The invention discloses a small sample training method based on target segmentation, and belongs to the field of image semantic segmentation. The method comprises the following steps of: 1, segmenting a target class of a small sample data set to generate an initial pixel-level label; 2, iteratively training a semantic segmentation network on the small sample data set to generate a fine pixel-level label; 3, acting a segmentation result on an original image to obtain a target image; 4, fusing the target image into a large number of images, generating a new image with the target image, and further generating a large number of new images with pixel-level labels; and 5, training the semantic segmentation network by using a large number of new images with pixel-level labels so as to train a reliable semantic segmentation network. According to the method, the time-consuming and labor-consuming operation of manually annotating the pixel-level semantic tag is solved, the refined segmentation of the target can be realized, and good annotation data is provided for image segmentation and target detection.

Description

technical field [0001] The invention belongs to the field of image semantic segmentation, in particular to a small-sample training method based on object segmentation. Background technique [0002] Image semantic segmentation is a task of computer vision. It assigns a semantic label to each pixel, that is, each pixel in the image has a label. The label of is 1, the label of the background is 0. Semantic segmentation is widely used in various fields, including: automatic driving, image diagnosis and other fields. [0003] Image semantic segmentation methods are mainly divided into two categories: strongly supervised semantic segmentation and weakly supervised semantic segmentation. Strongly supervised semantic segmentation requires manual annotation of a large number of pixel-level semantic labels, which is time-consuming and labor-intensive and the quality of the labels varies from person to person; weakly supervised methods do not require pixel-level image labels, but use...

AI Technical Summary

Problems solved by technology

[0004] The purpose of the present invention is to solve the problem of labeling a large number of pixel-level labels in strong supervised semantic segmentation and requiring a large number of training pictures in weak supervised semantic segmentation, and to provide a small sample training method based on target segmentation

Images