A Method for Learning Labeled Noisy Images Based on Dual Active Query

Abstract

The invention discloses a label noise image learning method based on double active query, which comprises the steps of: building a deep convolutional neural network classifier and a noise transfer matrix, and pre-training the classifier and the noise transfer matrix on a noisy data set; Select the image set and obtain the real label of the selected image set; use the selected image set and real label to update the noise transfer matrix; initialize the classifier based on the noise transfer matrix; construct the loss function on the real label and the loss function on the noise label; based on Stochastic gradient descent minimizes the true label loss and noise label loss, and updates the classifier parameters; repeats the iterative process to the maximum number of iterations K, and completes the training of the deep convolutional neural network classifier g. The invention uses the noise transfer matrix to establish a noise model, and introduces real labels to estimate the noise transfer matrix and improve the classifier; by designing a dual active query method, it saves labeling costs and maximizes the learning accuracy of the classifier.

Description

technical field [0001] The invention belongs to the technical field of image learning, and relates to a label-noisy image learning method based on double active query. Background technique [0002] The deep convolutional neural network image classifier model requires a large number of labels, but the actual collected labels often contain a lot of noise, and the deep neural network is easy to overfit the noisy data, which limits the generalization performance of the classifier model . [0003] Existing learning theory shows that by building a model for noise, if the noise model estimate is accurate, then the optimal classifier on noisy data is equivalent to the optimal classifier on real label data. [0004] The noise transfer matrix is ​​a commonly used noise model, which contains the flip probability between each category of the image, so as to realize the mapping between the real label probability distribution and the noise label probability distribution. Whereas estimat...

AI Technical Summary

Problems solved by technology

[0006] The purpose of the present invention is to propose a label-noisy image learning method based on dual active query to solve the problem that the existing deep neural network is easy to overfit the noise label and the noise transfer matrix estimation deviation caused by the traditional active query preference, Maximize the learning accuracy of the classifier while saving the labeling cost

Images