Scratch image data amplification method based on generative adversarial network

Abstract

The invention discloses a scratch image data amplification method based on a generative adversarial network. The scratch image data amplification method comprises the following steps: S1, constructinga scratch image data set; S2, preprocessing the scratch image; S3, constructing a generative adversarial network; S4, performing confrontation training; and S5, outputting a scratch image amplification data set. The main beneficial effects of the method are that a large amount of scratch image data can be generated by utilizing a generation network, effective characteristics of a real scratch image can be learned by using a discrimination network. Through performing the adversarial training on the generation network and the discrimination network, the probability distribution statistical lawdifference between the generated image and the real image can be effectively reduced, so that the requirement of a defect detection system based on deep learning on the number of training samples is met, and the quality requirement of the defect detection system on the training samples is also met.

Description

technical field [0001] The invention relates to an image data amplification method, in particular to a scratch image data amplification method based on a generative confrontation network, which belongs to the field of computer vision. Background technique [0002] In recent years, thanks to massive training samples and increasing computing power, deep learning has made breakthroughs in theoretical research and multiple application fields, especially in the field of computer vision. Deep learning simulates the neural connections and signal processing methods of the human brain, combines low-level features of high-dimensional data, and gradually abstracts more complex feature representations through hierarchical forms, but it also brings a huge number of parameters. In order to avoid overfitting of the model, a large number of samples are required to train the model, however, the cost of collecting and labeling samples is very high and time-consuming. Existing image data augm...

AI Technical Summary

Problems solved by technology

Deep learning simulates the neural connections and signal processing methods of the human brain, combines the low-level features of high-dimensional data, and gradually abstracts more complex feature representations through hierarchical forms, but it also brings a huge number of parameters

In order to avoid overfitting of the model, a large number of samples are required to train the model, however, the cost of collecting and labeling samples is very high and time-consuming

Existing image data augmentation techniques mainly include horizontal flipping, cropping, and contrast adjustment, etc. However, these traditional methods have a very limited amount of image data augmentation, and it is difficult to meet the requirements of deep learning-based application systems (such as defect detection systems). training needs

In addition, the number of surface defect images is less than that of natural images, and it is more difficult to collect and label

Images