Neural network-based automatic image annotation method, system, device and medium

Abstract

The invention relates to a neural network-based automatic image annotation method, a system, a device and a medium. The method comprises the steps of extracting image features of an experimental dataset by utilizing a pre-trained convolutional neural network model; according to the image features, calculating in a training set to obtain a neighborhood image set of the to-be-labeled image and a corresponding first label domain; constructing a label semantic association model between the first label domain and a second label domain corresponding to the training set, and performing calculation in the second label domain according to the label semantic association model to obtain a third label domain associated with each first label; calculating the similarity between the to-be-labeled imageand each neighborhood image, obtaining a first probability that each first label becomes a target label according to all the similarities, and obtaining a second probability that each third label becomes the target label according to all the first probabilities and the label semantic association model; and obtaining a target label according to all the similarities, all the first probabilities andall the second probabilities, and completing labeling according to the target label.

Description

technical field [0001] The present invention relates to the technical fields of computer vision and artificial intelligence, in particular to a neural network-based automatic image labeling method, system, device and medium. Background technique [0002] The automatic image annotation method is considered to be an effective solution to the semantic gap between the original image and its semantic information. It automatically learns the relationship between the semantic concept space and the visual feature space by using the training set images that have been marked with keywords. The latent correspondence or mapping model can then predict the semantic information of the unlabeled image through the constructed mapping model. [0003] Some existing methods use traditional machine learning algorithms to construct the mapping from semantic concept space to visual feature space, for example, by using the improved FCM clustering algorithm to divide different semantic image dataset...

AI Technical Summary

Problems solved by technology

[0004] However, the above-mentioned methods of using machine learning models to solve the gap between the original image and its semantic information have certain defects.

The label of the image close to the cluster center is selected by the clustering method, and the image label is realized by passing the label of the nearest neighbor image. Although these clustering and nearest neighbor methods can realize automatic image labeling, they only consider the image The relationship between the label and the image, without considering the relationship between the label and the label, and the lack of prediction of the number of labels, but in practice, the relationship between the labels is also a very important factor in predicting the semantic information of the image

Therefore, although the above two methods of automatic image annotation using clustering and nearest neighbor methods solve the gap between the original image and its semantic information to a certain extent, the accuracy of the annotation is not high, and the annotation effect is not good.

Images