Long-tail learning image classification and training method and device based on mixed batch normalization

Abstract

The invention discloses a mixed batch normalization-based long-tail learning image classification and training method and device, and the training method comprises the steps: obtaining a corresponding standardized feature vector through constructing a mixed standardized branch and a split standardized branch, inputting a sample image subjected to strong enhancement and weak enhancement into the standardized branch through a double-branch learning frame, and carrying out the recognition of the corresponding standardized feature vector. And carrying out image classification, calculating similarity maximization loss of classification prediction of the two branches through a classification result, calculating classification loss by using balanced cross entropy loss, and optimizing network parameters corresponding to the double-branch framework. The hybrid standardized branch can more comprehensively model a feature space and reduce the dominant position of the head class, and the split standardized branch can diversify the estimated Gaussian distribution, so that the Gaussian distribution can more comprehensively fit the training sample of the tail class, and the training efficiency is improved. According to the image classification method and device, the trained hybrid standardized branches are used for image classification.

Description

technical field [0001] The present invention relates to the field of image classification, in particular to a long-tail learning image classification, training method and device based on hybrid batch normalization. Background technique [0002] The imbalanced learning problem has attracted extensive research interest in recent years. However, existing methods cannot achieve high accuracy for tail classes without hindering the performance of head classes or maintaining an efficient frame. The goal of the present invention is to use long-tail training data for learning while avoiding the above problems. When using long-tail samples to learn deep convolutional neural networks (CNNs), the optimization of network parameters is dominated by the head class samples, resulting in relatively low performance of the tail class samples. In the field of image classification, it will eventually affect the performance of image classification. precision. The traditional solution to the da...

AI Technical Summary

Problems solved by technology

However, the impact of data imbalance on feature representation learning still cannot be effectively mitigated, as they still rely on data resampling or reweighting algorithms to manage multiple classifiers

Based on the above analysis, the existing deep neural network CNN still has no effective solution to the above-mentioned problems of classification and recognition for image datasets with long-tail features.

Images