Research On Airtightness Water Detection Method Of Airtight Container Based On Deep Separable Convolutional Neural Network

Airtight containers are widely used in people’s lives and industrial production.To ensure product quality and the safety of the users,factories need to carry out airtightness testing when producing airtight containers.This is why the study of airtightness testing methods is an important issue in the field of industrial production.The study of a convenient and efficient method for testing airtightness is of great relevance for improving production efficiency,improving production processes and reducing costs.Water detection method is one of the commonly used air tightness detection methods.The traditional artificial air-tightness water detection method can intuitively judge whether the container leaks,but it requires high observation time and labor cost,and cannot be used for batch detection.The subsequent research on the machine vision based air tight water detection method is limited by the limitations of traditional machine vision detection methods,the detection accuracy of leaking bubbles is not high,and their anti-interference performance is poor,so it is difficult to apply to actual detection.In recent years,with the rapid development of deep learning and the substantial improvement of the computing power of hardware devices,the advantages of convolutional neural networks in the field of object detection have gradually emerged.In this paper,the object detection method based on convolutional neural network is applied to the detection of leaking air bubbles,so as to obtain a convenient and efficient method for airtight water detection of airtight containers.This paper mainly does the following work:Firstly,an airtight water detection simulation device is built in this paper to generate leaking bubbles and collect leaking bubble images.The collected leaking bubble images are manually annotated and made into a dataset that meets the training requirements of the object detection model.Faster R-CNN target detection algorithm is selected for leak bubble detection.The detection effect is compared with traditional machine vision detection methods,which verifies the feasibility and effectiveness of the air tight water detection method based on convolution neural network.Secondly,this paper compares the performance of various target detection models,and finally chooses YOLOv5 s target detection model as the basic model for this research.To solve the problem that the bubbles are relatively concentrated and most of the bubbles are small in size and the color contour features are not obvious,this paper improves YOLOv5 s from the directions of increasing the attention mechanism,improving the feature fusion method,multi-scale detection,and optimizing the loss function.Compared with the base model YOLOv5 s,the average accuracy of the improved model B-YOLO in leaking bubble detection is improved by 3.8%.Finally,in order to improve the detection speed,reduce the pressure of model deployment and reduce the cost of enterprise equipment,this paper performs model compression on the B-YOLO model.Firstly,we design a lightweight network structure by taking advantage of the fact that the depth-separable convolution is less computationally intensive and less parametric than ordinary convolution.We then pruned the model by using a pruning technique and compared the pruning results to select a pruned model with a pruning rate of 0.6 as the final lightweight leaky bubble detection model.The final lightweight model not only achieves 2.5% better detection accuracy than YOLOv5 s on the leaking bubble dataset,but also increases the detection speed from 28 frames per second to47 frames per second.