Image depth clustering method and system based on self-supervised contrast learning

Abstract

According to the image deep clustering method and system based on self-supervised comparative learning, comparative learning is utilized to improve the discrimination of embedding, and under the condition of not giving human annotations, the comparative learning can learn the embedding with high cosine similarity and strong discrimination for semantically similar samples by discriminating the samples. On the basis, according to the technical scheme, the subtasks capable of simplifying the learning process are mined, and due to the fact that the intra-class difference of samples of the same class is smaller than that of samples of different classes, it is determined that the subtasks are the most natural division mode according to the classes of the samples. Therefore, compared with a mixedexpert system, highly professional experts are encouraged, each expert is good at processing samples of a specific category, and a good clustering result is naturally obtained. Meanwhile, compared with a hybrid expert system, a single objective function is optimized, clustering degradation can be prevented without processing such as pre-training or regular terms, and the method can be applied tounsupervised clustering tasks of more complex images.

Description

technical field [0001] The invention relates to the technical field of machine learning, in particular to an image depth clustering method and system based on self-supervised comparative learning. Background technique [0002] At present, unsupervised clustering is one of the challenging difficulties of deep neural networks. Since labeling a large amount of data requires high cost and intensive manpower, and the quality of labeling is difficult to guarantee, clustering is often required to divide the data into different Subsets can be used to mine the structural characteristics of the data set. [0003] At present, classical clustering methods such as k-means algorithm, spectral clustering and hierarchical clustering are not very effective when dealing with high-dimensional data. With the rapid development of deep learning, many studies have tried to combine classical clustering and deep learning methods, aiming to use deep neural networks to extract features, and then dire...

AI Technical Summary

Problems solved by technology

[0004] In the existing technology, the probability-based deep clustering method provides a systematic solution, which can prevent degenerate solutions concisely and effectively, but the existing methods limit the discriminability of the model

Methods based on deep generative models try to introduce a mixed Gaussian distribution prior to the latent representation of data to model multimodal data, but waste model performance in the process of learning data generation

However, discriminative clustering directly learns the mapping from input data to cluster labels, which is usually better than deep generation methods. Although the discriminative method has high computational efficiency, directly outputting low-dimensional cluster labels forces the model to discard the information of sample granularity, which limits Performance of feature learning and clustering

Existing probabilistic methods perform well on simple image datasets (such as MNIST), but perform poorly on more complex image datasets due to insufficient discriminative models

Images