Research On Key Technology Of Defect Drug Tablets Detection Based On Machine Vision

With the continuous improvement of people's requirements on the quality of drug production,it is particularly important to detect the defective drug tablets in the process of drug tablets production.At present,the detection of defective drugs by major domestic drug manufacturers can only be carried out on the drug pellets of blister type transparent packaging after the packaging of finished drugs.Other drugs,especially canned drugs,can only be estimated by manual sampling.Such an approach is inefficient,resulting in low reliability and high cost,making it difficult to adapt to the development of automation.With the continuous development of machine vision,machine learning and other theoretical research,a large number of industrial testing applications in the corresponding technology.In this paper,a prototype system of defect pill detection based on machine vision is designed,and the corresponding key technologies are studied to fill some gaps in the application of defect pill detection,so as to improve the degree of automation of pill detection.Aiming at the actual situation of filling tablet production,this paper adds image acquisition module and lighting system on the basis of the original granulator,and designs a defect tablet detection system based on machine vision with online detection function.The overall scheme includes the composition of various hardware subsystems,equipment selection and lighting scheme design.After analyzing the overall operation process of the system,the key steps of the detection system are divided into two stages,one is the online target image information acquisition stage,and the other is the detection and analysis of the collected target image to determine whether it is a defect pill.This paper studies the process of each stage of target acquisition,compares and analyzes the VIBE and PBAS algorithms in background subtraction,and improves the background subtraction algorithm according to the specific situation on this basis,which is more suitable for the system in general.Then the shadow of the acquired foreground target image is automatically removed by using the histogram inter-class variance method,and then the best target mask can be obtained.Aiming at the inevitable adhesion phenomenon in moving pill target,two adhesion segmentation methods based on morphology were designed.However,the whole algorithm process will reduce the frame rate of image processing.Based on the prior information of the system equipment,a multi-threaded scheme for image segmentation and a GPU acceleration processing scheme are designed to ensure the real-time performance of online target extraction.Finally,the principle of photoelectric switch is simulated,and the scheme of tablet counting using inter-frame information is designed.By studying the principle of perceptual hashing algorithm and using its rapidity of resolution on similar images,the principle of mean hashing and DCT hashing algorithm is compared and analyzed.After improvement,the perceptual hashing algorithm with more distinguishing ability for defective tablets is obtained.Support vector machine(SVM)and BP neural network(BP)are studied on the classifier to replace the simple hamming distance discrimination.On this basis,FasterRCNN network is adopted to detect and analyze the surface defects of tablets by using the deep convolution feature extraction structure of VGG16,and it is found that the features extracted by deep convolution are more distinguishable from surface defects.