Data sampling method and device

Abstract

The invention relates to the technical field of image recognition and provides a data sampling method and a data sampling device for solving the problem of long tail of a training data set in the prior art, and aims to solve the problem of incomplete network feature learning caused by the fact that training samples are the same in each network training process from the beginning to the end of network training, and the method comprises the following steps: acquiring the current iteration round of a network; updating the sample weight of each type of samples based on the current iteration roundof the network and the sample number of each type of samples, with the increase of the current iteration round of the network, the weight of the tail type of samples being gradually increased; and according to the updated sample weight of each type of samples, sampling samples meeting a preset condition as target samples. Through the increase of network iteration turns, the sample weights in the training data set are increased, but the increase amplitude of the weights of the tail class samples is large, the probability that the tail class samples are selected as the training samples is increased, the problem of long tails can be effectively relieved, and the network feature learning effect is good.

Description

technical field [0001] The present invention generally relates to the technical field of image recognition, and in particular relates to a data sampling method and device. Background technique [0002] Data sampling, that is, selecting some or all pictures from the training data set as training samples for image recognition network training. Most of the training data sets used in the current image recognition network training process have long-tail problems. The long tail problem is that a small number of class samples contain a large number of samples in the training data set, which are called head class samples, and most class samples contain a small number of samples in the training data set, called tail class samples. Data sampling is performed based on the training data set with long-tail problems. During the data sampling process, head-like samples are often sampled as training samples for image recognition network training. Therefore, the trained network tends to pre...

AI Technical Summary

Problems solved by technology

The above-mentioned data sampling method to solve the long-tail problem, from the beginning of the network training to the end of the network training, the sampled samples are the same in each network iteration process, so the training samples of the network training are the same, which leads to the incomplete feature learning of the network. low recognition accuracy

Images