| Although traditional anomaly detection methods have much higher accuracy and efficiency than manual inspection,problems of missed detection and false detection still exist.The main problem is that most anomaly detection algorithms are based on deep learning,which requires a large amount of marked data to train the network model of anomaly detection.However,in reality,it is difficult to find all types of defect data,and it is difficult to have a large number of marked abnormal data as data sets.In addition,the limitations of the network model itself require improvement of the anomaly detection algorithm.Based on the above problems,this paper,based on the network framework of reinforcement learning,uses reinforcement learning to train the network without a large number of well-marked training labels,but only needs to interact with the environment to obtain profit value,and then updates the advantages of agent parameters,so as to continuously optimize the network structure and improve the accuracy and efficiency of the model.A research method of tire X-ray image anomaly detection based on neural batch sampling was proposed.The specific work is as follows:(1)The original data of tire X-ray images were obtained from the tire manufacturing workshop,and after preliminary noise reduction and enhancement processing,the data set of tire X-ray images in Pascal VOC2007 format was generated through the calibration software independently developed by the laboratory,which provided a reliable guarantee for the followup network research and training.(2)The standard Encoder network is corrected and upgraded in structure,and the overall network is Spatial Variational Auto-encoder(SVAE)frame structure.It is proposed that the Encoder space uses larger feature maps as latent variables to capture Spatial information explicitly.This is achieved by allowing latent variables to be sampled from a matrix-variable normal(MVN)distribution with parameters calculated from the encoder network.In order to increase the dependence between locations on the latent feature map and reduce the number of parameters,a spatial SVAE distributed through low-rank MVN is further proposed.Experimental results show that the image processing speed and training speed of the proposed method are obviously better than the traditional encoder model.(3)Change the idea of image anomaly detection from image space anomaly detection to feature space anomaly detection.When stochastic gradient descent or other iterative learning algorithms are used to train the model on such unbalanced data,after each gradient update of the data sample,the loss value of the non-abnormal data decreases because the model sees more non-abnormal data in the training process,while the loss value of the abnormal data sample will fluctuate.Therefore,a neural batch sampler is introduced to regionalize the X-ray image information,maximize the feature difference between abnormal and non-abnormal regions.After FIFO cache and regular update,the discriminator is finally used to discriminate errors and locate errors.In this way,the tire X-ray image anomaly detection task can be accomplished efficiently and accurately.The network model designed in this paper performs best in comparison with four classical detection models. |
