A Triplet network learning method based on a semantic hierarchical structure

Abstract

The invention provides a Triplet network learning method based on a semantic hierarchical structure, and the method comprises the steps: constructing the semantic hierarchical structure, carrying outthe hierarchical Triplet sampling, carrying out the hierarchical Triplet network training, and carrying out the enhancement through employing a bilinear feature, thereby updating the parameters of a network. According to the method, the semantic knowledge is used for guiding the hierarchical distinguishing of the sample structure of the network, and the relation between the layers is used for enabling the network to pay more attention to more effective Triplets pairs, so that the effectiveness of samples in the batch is fully mined, and the separability of deep features learnt by the network is improved. Meanwhile, image details are enhanced by utilizing a bilinear function, and the image details and Triplet are jointly trained, so that the performance of the network is further improved.

Description

technical field [0001] The invention relates to the field of machine learning and image feature extraction, in particular to a Triplet network learning method. Background technique [0002] Metric learning is an efficient deep learning method that calculates the error by calculating the similarity between two images to update network parameters. The goal of metric learning is to calculate the similarity between pictures, so that the similarity of heterogeneous pictures is small, and the similarity of similar pictures is large, and finally learn the features with strong separability. From the original Siamese network to the current Triplet network and various improvements and variants of the Triplet network, the method of metric learning reflects its unique superiority. [0003] The Triplet network usually requires input of a triplet, including three samples of anchor, positive, and negative, where positive and anchor are from the same category, and negative and anchor are f...

AI Technical Summary

Problems solved by technology

[0005] But the Triplet network also has many shortcomings: for example, the loss function of the network converges very slowly; the sampling method has a great influence on the network, and whether the network can converge well depends largely on the selected triplet samples Is it reasonable? Moreover, the Triplet network is difficult to train on a large-scale data set. As the number of samples and categories increase, the network will generate a lot of Triplets that do not contribute to the training, and we often need to search the entire space to find the right network training. Really Contributing Hard Triplets

Although these methods showed good results, they did not take hard Triplets into account, so some studies combined the advantages of classification and hard Triplets to improve them. At the same time, in order to make full use of the batch, some studies also proposed to generate hard online in the batch. Triplets

Images