Research On Zipper Quality Inspection Algorithm Based On Combined Deep Neural Network

With the introduction of the Industry 4.0 strategy,intelligence and automation have become a new development trend in the manufacturing industry.As an important component of daily necessities,the zipper is an ideal practice object of intelligent manufacturing.In the actual production process,the appearance quality inspection of the finished zipper is the key to control the quality of the zipper.At present,the quality inspection of zippers mainly relies on manual visual inspection,which has the disadvantages of high cost,slow inspection speed and unstable inspection effect.The intelligent quality inspection algorithm can meet the requirements of high detection accuracy,fast detection speed and be able to deal with complex defects in actual industrial scenarios,and has practical application value.This research is based on the zipper image acquisition mechanism and deep learning technology,and it takes the plastic steel zipper as the research object to study the appearance quality inspection algorithm of the finished zipper.The zipper quality inspection system collects the grayscale image of the finished zipper section by section,and inputs each section of the image into the quality inspection algorithm for defect detection.The zipper can be divided into three parts by the similar structure of the zipper: tape,teeth and special parts.The special parts include the top stop and latch.This study carried out defect detection on the three parts of the zipper.The work content of this paper includes:(1)Aiming at the characteristics of complex zipper defects and similar zipper structures,a secondary detection mechanism is designed to locate zipper parts first and then detects the defects separately.Based on the secondary inspection mechanism,a multi-model scheduling quality inspection algorithm is proposed.The multi-model scheduling quality inspection algorithm uses the characteristics of the industrial zipper image to locate the zipper teeth,and uses the object detection model to locate special parts and detect the defects of each part.(2)Aiming at the requirements of high detection accuracy and fast detection speed,this paper compares the performance metrics of the two-stage object detection model Faster R-CNN and the single-stage object detection model ATSS,and the Reg DNet model combining the neural network Reg Net and the deformable convolution DCN v2 is proposed to further improve the detection accuracy of defects under the premise of satisfying real-time performance.(3)Aiming at the characteristics of small proportion of special parts and few occurrences,a combined quality inspection algorithm for double inspection in special parts is proposed based on the multi-model scheduling quality inspection algorithm.The combined quality inspection algorithm compares the classification effects of four lightweight classification models,and then selects the best classification model Squeeze Net to classify defects in special parts,thereby improving the recall rate of zipper defects.The experiment divides the collected zipper images into parts and cooperates with the annotations of the factory professionals to form the data set of all parts.The experimental results show that the m AP of the Reg DNet is higher than Faster R-CNN and ATSS,the inference time is only 56 ms,and the m AP value in all parts exceeds 0.88.The optimal classification model Squeeze Net has an average accuracy of 96.52% and 97.31% at the latch and top stop part,and the inference time is only 4.6ms.The defect recall rate of the quality inspection algorithm based on the pure detection model is 95.54%,and the defect recall rate of the quality inspection algorithm of the combined model is 96.87%.The inference time of the two quality inspection algorithms does not exceed 300 ms,which meets the requirements of industrial applications.The quality inspection algorithm proposed in this study quantifies the inspection results,solves the problem of manual inspection,significantly improves the production efficiency,and provides a basis for product industrial optimization.