Knowledge distillation data enhancement method and system based on generative adversarial network

Abstract

According to the knowledge distillation data enhancement method and system based on the generative adversarial network provided by the invention, through an easy-to-operate data enhancement mode, the positioning capability and the recognition accuracy of the convolutional network to the industrial data set are improved. According to the method, a generative adversarial network is combined to efficiently fit sample data distribution and an algorithm for improving the fault-tolerant rate of a neural network model in knowledge distillation, non-information noise is not introduced when an enhanced sample is generated, and the robustness of the model to an error tag is improved. According to the method, misleading on the model caused by marking errors of positive and negative samples of the industrial data set is reduced, the representation expression performance of the industrial data set on a semantic segmentation task is improved, and the sample feature learning ability of the model under the condition that feature granularity distinguishing is small is improved; The invention plays an important role in applications such as automobile part detection and manufacturing, railway part positioning and the like.

Description

technical field [0001] The invention belongs to the technical field of data enhancement, and in particular relates to a knowledge distillation data enhancement method and system based on generative confrontation networks. Background technique [0002] Data augmentation is a method of increasing the amount of data through a certain transformation method on the original data set. Its purpose is to have a sufficient proportion of samples when training a neural network classifier so that the classifier can adjust millions of parameters to the best point of model loss during the training process. [0003] The discarding, filling and fusion strategies of existing data enhancement often introduce more non-informative noise, reduce the signal-to-noise ratio of samples, and misidentify target regions during semantic segmentation. In addition, the difference between positive and negative samples of industrial data is small, and there are errors in the manual labeling process, and the...

AI Technical Summary

Problems solved by technology

[0003] The discarding, filling and fusion strategies of existing data enhancement often introduce more non-informative noise, reduce the signal-to-noise ratio of samples, and misidentify target regions during semantic segmentation.

In addition, the difference between positive and negative samples of industrial data is small, there are errors in the manual labeling process, and there are problems such as labeling area offset and errors.

These methods are unsatisfactory in terms of positioning ability and recognition accuracy

Images