Surface Defect Detection Algorithm Of Sanitary Products Based On Deep Neural Network

In recent years,as a basic industry for national economic development and the realization of the goal of strengthening the country,the manufacturing industry has increasingly high requirements for product quality.In the hygiene industry,surface defects are crucial to product quality assurance.However,the traditional manual visual inspection method has problems such as long time and high cost.Therefore,the researchers introduced deep learning to improve the automation level of surface defect detection of sanitary products.However,the deep learning method requires a large amount of labeled data for training in the surface defect detection of sanitary products,but it is difficult to obtain a large amount of defect data,high-quality labeled data is also scarce,and small sample detection is also facing difficulties.Therefore,this thesis aims to carry out research based on deep neural networks to solve the problems of lack of defect data,difficulty in labeling data,and difficulty in small sample detection in surface defect detection of sanitary products.The specific research content is as follows: In recent years,as a basic industry for national economic development and the realization of the goal of strengthening the country,the manufacturing industry has increasingly high requirements for product quality.In the hygiene industry,surface defects are crucial to product quality assurance.However,the traditional manual visual inspection method has problems such as long time and high cost.Therefore,the researchers introduced deep learning to improve the automation level of surface defect detection of sanitary products.However,the deep learning method requires a large amount of labeled data for training in the surface defect detection of sanitary products,but it is difficult to obtain a large amount of defect data,high-quality labeled data is also scarce,and small sample detection is also facing difficulties.Therefore,this thesis aims to carry out research based on deep neural networks to solve the problems of lack of defect data,difficulty in labeling data,and difficulty in small sample detection in surface defect detection of sanitary products.The specific research content is as follows:(1)A weakly supervised defect detection method combined with data generation is proposed.The method employs a two-stage defect detection model.Firstly,the problem of unbalanced distribution of different types of defect data is solved by designing a cut data generation strategy.This strategy introduces local irregularities in normal sample images by cutting arbitrary parts in normal samples and defect parts in defect samples,and pasting them on normal samples to generate fake samples.Then,a binary classifier is obtained by training normal samples and generated defect samples.In the second stage,defects are localized by learning and extracting features from the clipped patches.During the whole training process,only normal samples and a small number of defect samples are used,thus overcoming the problem of lacking a large amount of labeled defect data.Using this method,the method can not only simulate unseen defects in the real world,successfully solve the problem of insufficient samples and category imbalance in the sanitary product dataset,but also improve the detection performance of the model,significantly improving the detection of surface defects in sanitary products performance.(2)A small-sample hygiene product defect detection and classification method based on YOLOv5 is proposed.This method is a further study on the data set of sanitary products generated based on the data.It detects the defect target for the common defect types with position information on the surface of sanitary products to obtain the category information and position information of the defect.First,the YOLOv5 network model was selected as the basic defect detection model,and the YOLOv5 algorithm was improved for the problem of poor performance on small sample defects.This thesis proposes a new feature pyramid model CF-FPN and embeds it into PANet(Path Aggregation Network,PANet),so as to fully integrate the advantages of spatial and channel attention mechanisms.The design of the model can avoid information loss in the process of feature map generation,and can improve the representation ability of feature pyramid,thereby enhancing the accuracy of classification and regression prediction.Embedding CF-FPN into the YOLOv5 model can significantly improve the detection performance of the YOLOv5 model for multi-scale targets.The experimental results show that the detection accuracy of 97.5% is achieved on the sanitary product defect data set,especially the detection accuracy of small samples reaches 97.8%.Compared with several commonly used methods,this method is more superior in improving the accuracy of defect detection.