Research On Deformable Volume Visual Inspection Technology Based On Large Depth Of Field Multi-focal Surface Imaging

The rapid development of clinical medicine has also made infusion drugs widely used.However,in the production process,it is easy to be contaminated by impurities and it is difficult to find potential safety hazards.The labor intensity and unstable detection quality are the main disadvantages of the artificial light inspection.Machine vision detection technology can effectively solve the defects of traditional detection methods and can reduce production costs for long-term use.It is the preferred method in today's detection field.Firstly,this thesis briefly introduces the research background and significance of this subject,and the research and development process of visual inspection of liquid medicine was summarized.This thesis have discussed the limit of depth of field of the optical system which would could affect the accuracy of detection cause the fuzzy information.To slove this problem the relevant theories and technologies of image fusion was fully investigated include the new popular neural network-based fusion algorithm.Secondly,an experimental imaging system based on dual cameras to capture images with multiple focal planes was built in this thesis.An image fusion method based on convolutional neural network have been proposed.By treating fusion as a binary classification problem,the twin network model is used and designed,and Gaussian blur is used to simulate the out-of-focus image to create a multi-focus image data set pair model.It is used for training,and through the convolution of the fully connected layer,it is used for the fusion processing of multi-focus image pairs of different sizes in testing and practical applications.After testing the pre-trained CNN model,and by comparing the subjective and objective evaluation of different fusion methods,it can achieve a better fusion effect and achieve the effect of a large depth of field.Finally,after obtaining a sequence fusion image with a large depth of field effect and preprocessing such as difference,enhancement,and denoising of the image,a suspected moving target area is obtained,and image features such as its shape and gray scale are extracted to construct a feature vector.An AdaBoost classifier is construted and learned to classifies and identifies these suspected targets,several experimental exploration on the relevant properties of the classifier are conducted,and the classification algorithm parameters also are optimizesd.After compared with two other classifier(Decision Tree,SVM),the final classification accuracy of the AdaBoost classifier rate is 96 %,and The influencing factors of AdaBoost classification accuracy were also analyzed.