Ampoule Bottle Appearance Defect Detection Based On Adversarial Auto-Encoder And LBP Coding

Product defect detection is one of the important parts in the industrial production process,which ensures the quality and performance of products.It is difficult to collect large number of ampule samples with multiple defects in the actual production process,therefore the defect detection of ampule bottle appearance can be regarded as anomaly detection based on positive samples.This thesis summarizes the research status of anomaly detection,takes appearance defect detection of ampoules as the research topic,and combines the research results with industrial production application.The main work of this thesis is as follows:1.Research on the appearance defect detection model of ampoules based on adversarial learning and LBP coding.By modeling of gray image about defect-free ampoules based on convolutional auto-encoders,the defect detection of ampoule bottle was realized via adversarial learning and LBP coding.Based on the defect-free training samples and the known anormal samples,two encoding operations were carried out simultaneously to obtain the vector difference between the original image and the reconstructed image.Furthermore,by means of the optimal threshold estimation strategy based on weighted minimum error rate,the estimated optimal threshold can be used to detect appearance defect of ampoule bottles.At the same time,the heat map is used to display the specific area of the defect image.Finally,two-dimensional visualization based on T-SNE is used to demonstrate that the coding vector from the detection model has the ability of distinguishing between defect samples and defect-free samples.Taking the defect detection model based on original gray image data set as the baseline model,this thesis compares it with the model based on adversarial learning and LBP coding.The experimental results shows that the proposed defect detection algorithm can be helpful to improve the performance of defect detection.2.Implementation of the appearance defect detection system of ampoules based on adversarial learning.By combining the research results about ampule appearance defect detection system with the existing experimental platform,a system used for ampule appearance defect detection based on adversarial learning is designed.Based on the experimental platform of defect detection,this system carries out unified design and implementation of the logical framework,system module and software development of the appearance defect detection system of ampoule bottle,and completes the automatic monitoring process of defects.The defect detection results show that the system can accurately detect defects even though the proportion of positive and negative samples is out of balance.