Pedestrian re-identification method based on network regularization constraints based on easy-to-separable feature dropout

Abstract

The invention discloses a deep learning network regularization constraint method based on the discarding strategy of easy-to-separable features. At the same time, the difficult features are retained, and the network is restricted to improve the discrimination ability of features and the generalization of the network only by learning difficult features. The present invention applies the easy-to-separate feature discarding strategy to the fully connected layer and the convolutional layer, especially finds out the easily-separated rectangular area on the feature map through a sliding window method and sets it to zero, which solves the problem of the discrete state of zeroing. The network can automatically complete information according to the non-zero features on the feature map, which leads to the problem of weakening the effect of regularization, thus effectively constraining the training of the network and improving the generalization performance of the network.

Description

Technical field[0001]The present invention relates to depth learning and machine visual fields, and specific to the regularization constraints used in deep learning network training.[0002]technical background[0003]Due to the fact that the depth learning network training parameters are too many, the prefraction has been unavoidable when the network training has been trained. In response to the above problems, in addition to the data enhancement method in the network input, the more common method is to regularly process the network intermediate layer output. For example, Dropout and DropConnect methods are widely used regularization means, the former randomly randoms the output of each network node with a certain probability random, then randomly randomly randomly randomly randomly randomly randomly Zero. However, since the characteristics of the convolutional neural network output feature have highly correlated, when zero feature is too discrete, the network is easy to make informati...

AI Technical Summary

Problems solved by technology

However, since the features in the output feature map of the convolutional neural network are highly correlated, when the zero-setting feature is too discrete, the network is easy to complete information through non-zero features, which will weaken the effect of regularization. Therefore, the DropPath method further proposes a random pair The output of the network is set to zero and prevents other parallel networks from co-adapting, which greatly improves the independence of each sub-network, but setting the output of a sub-network to zero means that the ownership value of the network will no longer update and change, which will reduce The learning efficiency of the network, and this method can only be used for networks with fractal structures

[0004] It can be found that the above regularization constraint methods all adopt a random strategy and treat the output features equally. However, for the features extracted by the network, there are easy-to-separate features and difficult-to-separate features.

Images