A face noise dataset CNN training method based on the overall cosine distribution

Abstract

The invention proposes a face noise data set CNN training method based on overall cosine distribution, which belongs to the field of image recognition. This method first selects the face training data set containing noise and builds a benchmark convolutional neural network backbone; adds layers to the convolutional neural network backbone to obtain an updated convolutional neural network; obtains small Batch sample set, input each face picture sample in the small batch sample set into the updated network, get the corresponding cosine value and add the cosine distribution; calculate the corresponding training weight for each cosine value by designing three strategies and fusion functions , use the gradient descent algorithm to learn the updated network, and when the number of iterations reaches the upper limit, the training is terminated. The method has good robustness and can quickly and efficiently train noisy face training datasets without prior knowledge.

Description

technical field [0001] The invention relates to the field of image recognition. In particular, it relates to a face noise dataset CNN training method based on the overall cosine distribution. Background technique [0002] In recent years, deep learning technology has made rapid progress in both algorithms and hardware, and the scale of data it relies on has become larger and larger. Taking face recognition as an example, training a convolutional neural network for face recognition CNN requires large-scale training data sets. Although most of these training data sets are cleaned and filtered manually or by machines, there is still a certain proportion of noise. Face training data refers to face sample pictures and their corresponding labels, and face noise data refers to face sample pictures marked with labels that are not real labels. The IMDB-Face method mentioned that Microsoft's MS-Celeb-1M training data set contains nearly 50% noise data, and the depth model trained di...

AI Technical Summary

Problems solved by technology

[0013] 2.2) The dependence on hyperparameters is serious, and the convergence of the training process is difficult;

[0014] 2.3) For symmetrical intra-class noise, the recognition effect is relatively poor;

[0017] 3.1) This method needs to pre-determine the noise ratio of the data set. Usually, this prior knowledge is difficult to obtain, thus limiting its usage scenarios;

[0018] 3.2) Since the loss value of a single small batch sample set fluctuates greatly, it is not very accurate to filter noise data based on the small loss of a single small batch sample set;

[0019] 3.3) This method requires an isomorphic dual network structure, and the video memory usage and calculation consumption are basically twice that of the general network, which is not very practical for large networks with limited video memory;

Images