A Supervised Deep Hash Fast Image Retrieval Method and System

Abstract

The present invention proposes a method and system for fast image retrieval with supervised deep hashing. The method includes: constructing a deep convolutional neural network H″ for fast image retrieval; inputting the images in the gallery into the deep convolutional neural network H″ in sequence Get the real-valued features, get the hash code after the quantization operation and store it locally; input each query picture q into the deep convolutional neural network H″ and quantize to get the hash code h(q), and then calculate the hash code The Hamming distance between h(q) and all the hash codes stored locally, the Hamming distance is considered to be small as the high similarity, and it is sorted based on this, and finally the corresponding number of the most similar ones is returned according to the retrieval quantity requirements The picture. Based on the existing deep neural network, the present invention uses triplet label data to learn the image feature expression and uses triplet quantization loss function to build a supervised deep hash model, so as to achieve fast and accurate Image retrieval.

Description

technical field [0001] The invention relates to the fields of computer vision and image processing, in particular to a supervised deep hash fast image retrieval method and system based on a triple quantization loss function. Background technique [0002] With the rapid development of information technology, massive amounts of data are continuously generated, and the scale of image data is increasing exponentially. The huge amount of data makes direct retrieval of similar images bring great time and space overhead. Therefore, how to quickly retrieve similar images from massive images has become an urgent problem to be solved. Hashing has become a common solution as a way to map images into low-dimensional binary codes. In recent years, the deep convolutional neural network has developed rapidly, and the deep hashing method based on it has shown great potential in the field of fast image retrieval. In particular, supervised deep hashing methods have received extensive attent...

AI Technical Summary

Problems solved by technology

[0004] However, the defect of this method is that the loss of semantic information caused by reducing quantization is not equivalent to learning approximately binarized image features; on the contrary, this is a very strong constraint for network learning tasks, making the learned features themselves Only contains very little semantic information, in other words, the quantization loss function designed by Liu et al. leads to a suboptimal solution

Images