Online learning-based potential semantic cross-media hash retrieval method

Abstract

The invention discloses an online learning-based potential semantic cross-media hash retrieval method, which realizes cross-media retrieval of image and text modes. The method comprises the followingsteps of establishing an image and text pair data set, extracting features of data in the data set, performing mean removal, and dividing the data set into a training set and a test set according to acertain ratio; mapping discrete tags to continuous potential semantic spaces, and building an objective function by utilizing the similarity between the retention data; solving the objective functionby utilizing an online learning-based iterative optimization scheme, and when new data is generated, updating a hash function by only utilizing the new data, thereby improving the efficiency of a training process; and calculating hash codes of the image and text data in the test set by utilizing the hash function, by taking the data in one mode in the test set as a query set and the data in the other mode as a target data set, calculating Hamming distances between the data in the data query set and all the data in the target data set, performing sorting according to an ascending order, and returning the heterogeneous data sorted in front to serve as cross-media retrieval results.

Description

technical field [0001] The invention relates to the fields of multimedia retrieval and pattern recognition, in particular to a latent semantic cross-media hash retrieval method based on online learning. Background technique [0002] In recent years, due to the high efficiency and effectiveness of hashing methods on large-scale data sets, researchers have attracted extensive attention; the goal of hashing methods is to integrate data into Mapped to Hamming space; the similarity between data can be efficiently calculated by XOR operation, which greatly speeds up the retrieval speed under the premise of ensuring retrieval performance; however, most hash methods are only for the application of a single modality. With the rapid development of digital devices, there are more and more multimedia data on the network; data of different modalities can represent the same semantic data, which limits the application of single-mode hashing methods; data, but it is expected to return simi...

AI Technical Summary

Problems solved by technology

[0002] In recent years, due to the high efficiency and effectiveness of hashing methods on large-scale data sets, researchers have attracted extensive attention; the goal of hashing methods is to integrate data into Mapped to Hamming space; the similarity between data can be efficiently calculated by XOR operation, which greatly speeds up the retrieval speed under the premise of ensuring retrieval performance; however, most hash methods are only for the application of a single modality. With the rapid development of digital devices, there are more and more multimedia data on the network; data of different modalities can represent the same semantic data, which limits the application of single-mode hashing methods; data, but it is expected to return similar data of various modalities; however, the similarity between heterogeneous data cannot be directly measured, and how to measure the similarity of heterogeneous data becomes a challenge. The cross-media hashing method maps heterogeneous data to a shared Hamming space where the similarity of spatially heterogeneous data can be computed efficiently

[0003] Recently, researchers have proposed a variety of cross-media hashing methods and achieved satisfactory results; it has been proved that using the supervised information of the data (such as category labels) can generate hash codes based on high-level semantic preservation. Improve retrieval performance; however, discrete labels cannot accurately measure the similarity between data, which will lead to a decline in the ability of hash functions to distinguish; in addition, although research on cross-media hashing has made some progress, most existing methods are based on batch data This type of method requires all training data to be available before learning the hash function, but in practical applications, multimedia data on the network will continue to be generated over time, for example, billions of images are uploaded to Internet; when new data is generated, these methods have to use all the accumulated training data to retrain the hash function; this makes the hash method lose its efficiency, especially when new data is frequently generated; in addition, with the new data Continuously generated, the training data set becomes very large; on the one hand, the training data occupies too much memory, which makes it impossible to load all the data into the memory at one time; on the other hand, even if the memory is sufficient, the training time is usually unacceptable; In order to solve the above problems, the present invention proposes a cross-media hash body retrieval method based on online learning, which uses discrete labels to learn a continuous latent semantic space to more accurately measure the similarity between data, so that the returned retrieval The result is more accurate; and this method effectively realizes that when new data is generated, only the new data is used to update the training hash function, making the training of the hash function more efficient and reducing memory overhead

Images