Research On Optic Inspection Technique For Pharmaceutical Infusion Bottles

Medical infusion is one of the five major medical preparations in China,and it is widely used in clinical practice.Obviously,the quality of the infusion bottle packaging directly affects the safety of the patient.However,due to the immature manufacturing process of the infusion bottle,some appearance defects will show up stably with a low probability.At present,in order to ensure the packaging of the infusion bottles produced,the pharmaceutical factory adopts the method of manual detection.However,manual detection has the problems of inefficiency and insufficient reliability.Therefore,the thesis studies the visual inspection technology of the appearance defect of medical infusion,and cooperates with hardware acquisition equipment and processing equipment to achieve the purpose of replacing manual detection.Improve efficiency and precision,reduce labor costs,and have enormous economic and social benefits.Currently,the number of papers on the appearance of infusion bottles is inadequate,and the quality is not good,while most of them concentrate on a small part of it.Rare people have proposed a whole plan.Therefore,this thesis proposes a complete solution to several common defects,including image preprocessing,image segmentation and localization,and image post-processing(image classification),with an in-depth discussion of each part.For the preprocessing part of the image,the thesis discusses the image alignment algorithm to solve the image jitter caused by hardware.Firstly,the specific source and cause of the picture jitter are analyzed,and the specific presentation form on the image is determined.Starting from three different ideas,the method based on feature matching,the method based on template and gray scale and the method based on correlation coefficient are used to solve the image registration problem and optimized separately.Through experimental comparison,the best image registration algorithm in this scenario is determined,which is an improved template-based method.The algorithm achieves the second highest precision and has a speed far exceeding 5 times that of other algorithms,which satisfies the actual needs.Aiming at the segmentation and localization part of the image,the thesis discusses the high-performance ellipse detection algorithm that meets the industrial needs,in order to fully utilize the ellipse that frequently appears in the infusion bottle image.Firstly,starting from the simple circular detection,from the traditional method of line segment approximation,the optimization based on the arc support line is designed,and the polarity limitation and regional limitation of the pair of lines are designed.After success in the circular detection,we expands to the elliptical detection,upgrades the arc support line segment,and designs a cascaded clustering method to solve the high-level clustering problem.By comparing with the other two classical elliptical detection methods RHT method and ELSDc method,the algorithm achieves the best results.For the post-processing part of the image,this paper discusses the use of lightweight network for image classification.This paper analyzes the requirements and characteristics of the detection scene and determines the route to use the small model for classification.Firstly,the idea and network structure of lightweight network are studied.The calculation cost and parameter distribution are counted,and the optimization of 1 x 1 pointwise convolution using optimization function is used.To further reduce network latency and flexibly adjust the size of the model,a breadth factor and a resolution factor are used.In the experimental comparison,the improved lightweight network is slightly lower in accuracy than the classic VGG-16,but the speed is far superior;compared with the traditional algorithm,the speed is slightly inferior,but it performs better for some scenes.