Zero-sample image hash retrieval method based on visual-to-semantic network

Abstract

The invention discloses a zero-sample image hash retrieval method based on a visual-to-semantic network, and the method comprises the steps: converting an image feature vector into a semantic featurebased on the visual-to-semantic network, forming a target loss function by adopting the distance between the semantic feature and a class vector semantic feature, and achieving the classification lossand hash loss of semantic feature reconstruction. The optimal hash method on the training data set is jointly solved, so that the hash retrieval method still has feasibility for non-appeared categories, has better generalization and robustness compared with the traditional hash algorithm, reduces the training overhead of the image retrieval model at present when the data types become richer and richer, and slows down the update period of the hash model. The method can be applied to the fields of image retrieval, image tracing and the like in which database pictures are continuously expanded.

Description

technical field [0001] The invention belongs to a zero-sample image hash retrieval technology, in particular to a zero-sample image hash retrieval method based on a visual-to-semantic network. Background technique [0002] Image hash retrieval is an important problem in the field of computer vision. Its task is to quickly find the most similar image in the database based on the input image, sort it according to the similarity, and use the image correlation in the database to search for images of the same style. Or trace the source of the input image. With the increasing amount of data today, the traditional way of searching for pictures will cause the problem of too slow search speed due to too many pictures in the database, which cannot meet the ever-expanding demand for image retrieval. At the same time, the sharp increase in the amount of image data will inevitably bring about the enrichment of image types. At this time, the image hashing method using supervised learning...

AI Technical Summary

Problems solved by technology

There is a flaw in this approach. There is a degree of similarity between categories. For categories that are closer in the semantic space, their distance in the binary space, that is, the Hamming space, may be closer, which means that the learned hash mapping function Unable to distinguish between unseen categories of image data and similar categories of image data, this defect will directly lead to poor performance of hashing algorithms on unseen categories with similar categories

The two-stage algorithm first integrates image information and semantic information into the same attribute space, and uses attribute similarity and inter-modal similarity to generate hash codes. The two-stage algorithm can artificially enlarge the distance between categories to avoid the single-stage algorithm. The representative algorithm is AgNet [Z.Ji, Y.X.Sun, Y.L.Yu and Y.Gao, "Attribute-guided network for cross-modal zero-shot hashing," IEEE Transactions on Neural Networks and Learning Systems, vol 31, pp .321-330, 2020], although AgNet pays more attention to amplifying the gap between categories, its first step of attribute transfer learning has been proved to be poor in zero-shot learning.

Images