PET Bottle Appearance Inspection System Based On Machine Vision

With the development of machine vision,Computer intelligent learning become the future trend of industrial automation and intelligent Machine vision system with its unique advantages,gradually replaced artificial vision,has been widely used in industrial detection.The development of modern pipeline greatly stimulate the edible oil and beverage industry.Each line of output can reach tens of thousands of bottles per hour.In order to ensure the appearance quality of PET bottles.The traditional manual detection method have limits in heavy workload,high failure rate,lack of stability.So,the paper design and implement the PET bottle on-line detection system based on machine vision.The detection system include bottle cap detection,liquid level detection,printing detection and label detection.This paper mainly discuss the bottle cap detection and label detection.Based on the experiment,the system has used in practical application.The paper do the research about the key technology of the detection system:According to the production environment and the characteristics of the detection area,choose the standard of hardware equipment,and build this system.In the bottle cap detection module,choose the backlight,highlight the region contour of cap.In the label detection module,choose the linear light,capture a clear label for image matching.The bottle mainly divided into normal,low,high cover and askew cover.This paper improved the detection based on cap fitting with straight line detection algorithm,shorten the detection time.According to the environment,the bottle cap will gather water droplet In this case,the algorithm has a large error.For this reason,the paper presents the bottle detection algorithm based on corner detection.Through the endpoints of support ring and the points under sealing strip,as state judgement basis.Compared to the previous algorithms,this algorithm has better adaptability.The traditional feature selection algorithm based on multi sample analysis,but in practical application,the manual operation can only provide a few training samples,unable to obtain prior knowledge,select online feature does not apply to less sample.The paper improved the adaptive feature selection algorithm.According to FISHER criterion,proposed the evaluation characteristics of the evaluation standard,sorting the color feature and texture feature.The different categories of labeling samples obtain appropriate feature subset.Through feature matching,detect the label content.Through the experiments on real data sets,label detection accuracy rate reached 99.19%.