Research On Multii-Task Defect Detection Technology Of Beverage Bottle Based On Machine Vision

Beverage bottle factory appearance inspection systems play an important role in the entire beverage canning automation line.According to the order of the beverage filling line,the inspection items generally include label inspection,spray code inspection,liquid level inspection and high and low cap inspection,these inspection items have a vital role in factory quality control.This means that several different independent inspection systems have to be installed at different locations on a production line,which is a waste of manufacturing resources and difficult to install.It is also not easy to interconnect data with the database in this way,which affects product quality tracking and management.In this paper,the following research has been conducted to address the above issues in conjunction with computational vision techniques:Ⅰ.In response to the problems of the current factory routine inspection system for beverage filling lines,the overall architecture of the joint inspection system for multiple types of defects was completed according to the production requirements of beverage filling lines,specifically including detection,rejection,visual interaction and other functions.Ⅱ.Image pre-processing and data enhancement designs were carried out on the acquired images and datasets.The label and print images are pre-processed using methods such as image sharpening and enhancement to address vignetting and blurring of the captured images.In response to the insufficient number of data sets collected for network training,data enhancement is achieved by scaling,panning,rotating,mirroring,adding noise,and recombining colour channels for complex combinatorial shape-based labels,and Poisson fusionbased data enhancement is proposed for laser engraved printouts,which effectively expands the data set.Ⅲ.A neural network-based scalable defect detection network is designed to characterise the multiple types of defects corresponding to different processes in the beverage filling line.The network enables the direct identification of missing,broken and incorrect defects in beverage coding,as well as missing and backwards label defects,and provides a feature base for subsequent label breakage detection.The network also provides an expandable interface for adding and removing detection modules according to the actual task.In addition,to improve the speed of the designed network detection,a balanced improvement in speed and accuracy was made to the network by using a model pruning method.Ⅳ.Breakage detection of the label segmented images extracted by the joint multi-defect detection network.In this paper,a complete label region extraction algorithm based on probabilistic Hough transform and convex packet detection is designed to extract the complete region of the label.After obtaining the complete tag area,the tag breakage detection algorithm based on the colour space and Euclidean distance designed in the paper is then used to finalise the tag breakage detection.Ⅴ.Experimental system construction and algorithm functional testing.In this paper,we have completed the selection of important hardware for the experimental system and built the basic experimental platform.The experimental platform consists of an optical image acquisition module,a substandard reject module and a software control module.Afterwards,the designed inspection system was tested experimentally on a real platform for accuracy and speed.Experiments have shown that the joint multi-defect detection method proposed in this paper has a detection accuracy of more than 95%,and is able to improve the efficiency of defect detection and reduce the waste of resources.In addition,the system can be combined with database technology to facilitate the network management of production line faults,which helps to realise the integrated inspection and management of beverage production lines.