Unsupervised cross-modal hash retrieval method based on noisy label learning

Abstract

The invention discloses an unsupervised cross-modal hash retrieval method based on noisy tag learning. The method comprises: a multi-modal data acquisition step, a dual multi-modal neural network construction step, a similarity pseudo label with noise extracted based on a pre-trained neural network, an unsupervised cross-modal Hash training step based on noise label learning and a final dual cross-modal Hash retrieval test step. According to the method, two dual Hash model groups are set and relatively clean pseudo labels are mutually fed to each other, so that the learning of the model is misguided by noise pseudo labels as little as possible, the final effect is that the model is converged to a better position, and the performance on a test data set is better than that of other unsupervised cross-modal Hash methods.

Description

technical field [0001] The invention relates to an unsupervised cross-modal hash retrieval method based on noisy label learning, which is used for image hash retrieval with text description. Background technique [0002] The general cross-modal hash retrieval task is implemented through a set of networks, including image networks and text networks. In the case of supervised labeling, each image / text is assigned multiple labels. If a picture and a text have at least one common label among their respective labels, they are called similar labels, otherwise they are called not similar. In the process of training this group of networks, a batch of "picture-text" pairs is sampled each time, and its feature representation is obtained after network processing. The similarity is calculated based on the feature representation, and the error function of the real similarity is calculated. The error function is propagated backwards, so that the network weights are updated in the direct...

AI Technical Summary

Problems solved by technology

[0005] Purpose of the invention: Aiming at the problem that the previous methods generally use the picture features extracted by the pre-trained network and the text features of the bag-of-words model to calculate the cosine similarity as a pseudo-label to directly fit, and the obtained pseudo-label contains a lot of noise. The present invention An unsupervised cross-modal hash retrieval method based on noisy label learning is provided. In the present invention, the idea of ​​learning with noisy labels is used ingeniously, and two sets of dual hash models are set, respectively using the "small error criterion "To select relatively clean and credible similarity labels in each small batch, feed them to the other party's neural network, and do training. Practice has proved that it can effectively improve the generalization performance of the model.

Images