Tire Simulative Defect Image Generation Based On Generative Adversarial Network

Tn the production process of tires,most of the tire quality inspection work is done manually.At present,deep learning models have been used in many jobs to improve this situation.The effectiveness of deep learning largely depends on the diversity of samples.For tire defect samples,not all types of defect samples have sufficient diversity.Therefore,how to effectively increase the diversity of scarce samples is crucial to the application of deep learning models in tire defect detection.In this thesis,through intensive research on tire simulative defect image generation algorithm based on generative adversarial network,a controllable generation algorithm for tire simulative defect image based on specified background is proposed.The main research work includes the following parts:1.A basic model of tire simulative defect image generation based on the generative adversarial network,named Base Tire Defect GAN(BTD-GAN),is constructed.Aiming at the problems of model collapse and training instability that appears in the model,the objective function is optimized.Finally,the Progressive Grow training algorithm is used to enable the model to generate higher quality tire simulative flaw images with complex textures.2.Aiming at the problem that the tire defect category and shape cannot be controlled in the basic model,the research on the hidden space decomposition and feature decoupling algorithm of the generative adversarial network is introduced,and a network named Variational Tire Defect GAN(VTD-GAN)is designed.By adding an auxiliary classifier to the model,the tire defect categories can be controlled.Based on the idea of maximizing mutual information,continuous control variables are added to the model to control the shape of tire simulative defects.Finally,the two improvements are combined into a model to realize the image generation algorithm with controllable tire defect category and adjustable shape.3.Considering the practical applicability of defect image generation,a Background-Driven Variational Tire Defect GAN(BDVTD-GAN)is proposed.The main idea of this scheme is thar the tire simulative defect image generation algorithm is based on a given background.By adding the encoder structure to the model,the given background image information is extracted into a feature vector,and the latent variable z is synthesized through the redesigned latent variable synthesis operation,so that the defect image of the specified background can be obtained by generator.