Multi-scale automatic labeling method for images

Abstract

The invention relates to a multi-scale automatic labeling method for images. The method includes the steps of searching a training set for K-nearest images of images to be labeled; adopting the occurrence frequency of global labels in the K-nearest images as the first weight; selecting M K-nearest images with the highest matching degree to be added into a candidate set; according to the occurrencefrequency of each global label in the candidate set images, updating the first weight to obtain second weight; utilizing the candidate set images to conduct local labeling on the images to be labeled; calculating an average relevancy coefficient between each local label of the images to be labeled in the training set images and all the global labels of the training set images, conducting weightedsummation on the first weight and the second weight to obtain third weight, and obtaining t global labels with the maximum third weight as the global labels of the images to be labeled. By using themulti-scale automatic labeling method for the images, multi-scale labeling of global and local semanteme of the images can be achieved. During global labeling, relevancy information between the locallabels and the global labels is utilized, and the accuracy of the global labels is improved.

Description

technical field [0001] The invention relates to the technical field of machine learning, in particular to an image automatic labeling technology. Background technique [0002] Automatic image annotation technology is one of the research hotspots in machine learning and computer vision. Automatic image annotation is to automatically assign semantic information to images in the form of words and vocabulary. The automatic labeling of images has great application value in many fields. The basic idea is to train a model using the already labeled images, and then apply the model to the unlabeled images to be labeled, and then infer the image to be labeled. semantics. [0003] The semantics of an image can be divided into two scales: local semantics and global semantics. Local semantics refers to the semantics presented by a certain part of the image, which can establish a mapping with pixels at certain positions in the image; global semantics refers to the semantics reflected by ...

AI Technical Summary

Problems solved by technology

[0004] In the existing technology, semantic segmentation is an effective method for automatic annotation of local semantics of images. It realizes local semantics annotation by giving each pixel in the image a semantic label, and does not establish pixel and semantic correspondence. Semantic segmentation can more accurately realize local semantic annotation, but cannot achieve global semantic annotation.

At present, there is still a lack of research on methods for multi-scale semantic annotation based on semantic segmentation

Images