A method for identifying long-tailed distributions with two branches and multiple centers

Abstract

The invention belongs to the field of computer vision, and proposes a double-branch multi-center long-tail distribution recognition method for solving problems caused by long-tail distribution data sets. Input the image into the default branch and the resampling branch for data enhancement, and then input it into the deep convolutional neural network to obtain the low-dimensional feature representation; then pass through the fully connected layer to obtain the probability of belonging to each category, multiply it by a matrix representing the multi-center to obtain the feature matrix and take The maximum value, to get the probability of finally belonging to each category; calculate the loss separately; add up to get the final loss, backpropagate the network according to the loss and update the weight; iterate continuously; when the recognition task is required, input the picture again Sampling branch to get the probability that the picture belongs to each category. The impact of data distribution changes brought about by resampling can be mitigated by the dual-branch multi-center, and the impact of long-tail distribution can be further processed to bring better recognition and classification results, and the model has better generalization ability.

Description

technical field [0001] The invention relates to the field of computer vision, in particular to a double-branch multi-center long-tail distribution recognition method. Background technique [0002] With the rapid development of deep convolutional neural networks, the effect of image classification has achieved amazing results. This achievement is inseparable from increasingly rich data sets. Academically, the distribution of the number of category labels in most data sets is almost uniform, but the data in the real world is uneven, and even shows a long-tail distribution, that is, a small number of categories occupy most of the number of pictures, This part of the category is called the head category, and the remaining categories only occupy a small number of pictures. This part of the category is called the tail category. For details, see figure 1 . [0003] Existing popular methods for dealing with long-tailed distributions include resampling and reweighting. The essenc...

AI Technical Summary

Problems solved by technology

Although this can alleviate the problems caused by the long-tail distribution, it will also cause another problem

Images