Research On Anomaly Detection Algorithm Of Tire Defect Based On X-ray Image Reconstruction Residual

Defect inspection by X-ray image is an important procedure during the manufacture of tires.Presently,it relies mainly on manual check of X-ray images to judge whether there are defects in tires in China.In this paper,anomaly detection algorithms for tire defect inspection is proposed by using machine learning and deep learning.In industrial scene,it’s usual that we can only acquire a small number of defect samples but a large number of normal samples.Classical object detection algorithms usually require a large number of defect samples to train model,so it is difficult to apply these algorithms in our case.Therefore,we propose an anomaly detection approach based on image reconstruction and residual calculationFirstly,in this paper,we deal with the defect detection problem by using Principal component analysis(PCA)algorithm and wavelet filtering algorithm,and propose a novel method called Wavelet-PCA,which adopts PCA reconstruction to determine the threshold of wavelet filtering and employs a new threshold function.Afterwards,the reconstructed image and the original image are used for residual calculation.The model achieves up to 75.3%accuracy,65.8%recall,and 0.803 AUC.Secondly,in this paper,we design a multi-channel autoencoder for image reconstruction.The autoencoder only uses defect-free images during training phase to learn feature of defect-free images in depth,and the multi-channel structure can better capture the features.After the reconstruction,Siamese neural network was used to estimate the difference between the original image and the reconstructed image,and the feature vectors were used to assist classification.Owing to fewer parameters of Siamese neural network.We use a small quantity of defect samples to train the network,while the larger part of defect samples are reserved for testing resembling to the real industrial scene.Consequently,the model achieves 85.5%accuracy,89.9%recall,and 0.932 AUC.What’s more,the recall of defect detection with limited samples is up close to those with abundant samples,proving that our model has good performance on various types of defects.Finally,in this paper,we design an image reconstruction model called FMD-GAN(Feature module-driven defect detection Generative Adversarial Networks),which is able to reconstruct clearer images by introducing the idea of generative adversarial networks.At the same time,we designed a Momentum Update-based Siamese neural network(MUSNN)model to calculate the difference between the reconstructed image and the original image directly.This model not only augments the training samples,but also utilizes normal images more effectively,resulting in reduction of the misjudgment of normal images.In the end,we achieve 94.7%accuracy,96.8%recall,and 0.991 AUC.Moreover,the model is insensitive to the quantity of defect samples,and achieves high recall for various types of defects.