Research On Defect Detection Algorithm Of Flexible Packaging Tissue Paper Based On Machine Vision

Tissue paper is one of the most common wrapping papers in our daily life,among which tissue paper in the form of flexible packaging heat sealed with plastic film has been widely sold and used because of its portability and price advantage.Due to the increasingly fierce market competition and the strong demand of industrial upgrading for better quality and efficiency,the factory quality requirements are getting increasingly higher for flexible packaging tissue paper.However,manual visual inspection is still the main inspection method in the domestic industry at present,which leads to certain problems such as low efficiency,high false inspection and missed inspection,as well as absence of uniform inspection standard.In order to achieve automatic defect inspection in packaging production of tissue paper,this paper designs a defect inspection method for flexible packaging of tissue paper based on the actual production needs,which combines deep learning anomaly detection with traditional machine vision detection method.We collected the data set from the production line,and used it to verify the effectiveness and practicability of the inspection algorithm.According to the working order,it can mainly divide this paper into the following parts:Firstly,the research field of this project is the application of machine vision in packaging quality inspection industry.Therefore,the defect inspection technology based on machine vision and the application of machine vision in packaging industry at home and abroad are investigated in order to provide references.Then this paper conduct field investigation on the production line,analyzes the technological process of tissue paper flexible packaging,concludes the types of defects that may be caused in the production process,and then analyzes the characteristics of different defects to put forward appropriate imaging lighting schemes and algorithm design ideas.On the basis of analysis,this paper intends to adopt a fusion algorithm integrating deep learning anomaly detection and traditional defect inspection method.Secondly,this paper investigates the image anomaly detection algorithms commonly used in the field of industrial vision,summarizes several types of image anomaly detection algorithms according to the principle,introduces a benchmark data set for industrial image anomaly detection,and analyzes the advantages,applicable scenarios and relative disadvantages of different algorithms according to the performance of different types of algorithms on the data set,which provides a reference and inspiration for the establishment of anomaly detection algorithms appropriate for defect inspection of tissue paper flexible packaging in this paper.Thirdly,according to the previous investigation and research results,this paper designs an image anomaly detection algorithm for the application scenario of this project.First of all,the anomaly detection network architecture FCDD based on full convolutional neural network(FCN)and support vector description(SVDD)is analyzed,to study its operating principle,and then improves and quotes it.The convolutional layer of Res Net50 network is used to replace the convolutional part of FCDD network to achieve stronger feature extraction ability.The receptive field size and Gaussian kernel size of FCDD network are adjusted according to the characteristics of the project data to apply to the scene in this paper.Fourthly,a defect inspection algorithm for tissue paper flexible packaging is designed,to perform the anomaly detection on the input image.At the same time,the packaging foreground is extracted by image threshold segmentation based on edge information.The extracted packaging foreground is sent to the shape defect inspection algorithm module to extract the shape defects.Then the packaging foreground is compared with the heat map of abnormal area output by the anomaly detection algorithm module,to obtain the defect region image in the foreground range to classify defects.Defect classification is a multi-classifier based on MLP,which uses Mobilenetv2 as a feature extraction network.The extracted features are sent to MLP after PCA dimension reduction.At the same time,rejection classes are added during MLP training,so that MLP can classify images that do not belong to the training set category,i.e.,to identify untrained new defect types or anomaly detection misjudgment area images to a certain extent.Fifthly,according to the analysis in this paper,the imaging lighting scheme is designed,and the experimental hardware platform is built according to this scheme,and deployed in the production line for verification.Through the experimental platform,a large number of images are continuously collected to construct an image data set.According to the actual production,the performance of the algorithm on the data set is tested by the indicators of missed inspection rate,false inspection rate and defect classification accuracy rate,which verifies the effectiveness and practicability of the algorithm and concludes that the algorithm can meet the actual needs of industrial production.