Research On Optic Inspection Algorithm For Large Infusion Bottles

In recent years,because of the application of automatic filling technology,the annual production of infusion filling preparation is very large.The inspection of product quality is still in the stage of artificial visual inspection,which greatly limits the production efficiency.By using machine vision technology mainly based on deep learning algorithm,this project designs text detection system of infusion bottle label,which can replace manual detection and improve production efficiency.This topic to industrial medical infusion bottle production environment as the research background,around the infusion bottle label text detection and recognition method of appearance,with deep learning technology as the core,the development on the basis of forefathers' research,large infusion appearance text detection system is designed.This paper introduces the system requirement and composition,and verifies the stability and real time of the system.In this paper,a text correction method based on Fourier transform and hoff transform is designed to solve the problem of label text skew.By means of the Fourier transform of the input,the Angle detection and rotation correction of the image in the frequency domain are carried out.Finally,the method is tested and evaluated in terms of detection accuracy and detection time.According to the requirement of text detection,the network structure of convolution neural network and recursive neural network is designed.By using the convolution neural network and recursive neural network to form a combined network,the context information of the local area in the image can be effectively used to reduce the error detection and omission detection in the text area.This paper also designs a detailed text area method to improve the accuracy of text area level detection.This paper also designs a method to add trainable parameters to the network to refine the edge of the predicted text box,so as to further improve the fitting degree with the tag truth value area.In order to meet the needs of text recognition,a text recognition network based on associative timing classification is designed.First of all,this paper USES the recursive neural network structure based on the associative time series classification and the convolution neural network to train together,avoiding the operation of pre-segmentation and post-processing of text.Then through the expansion of data set,the overall effect of the network is improved.To test whether the system meet the demand of industry,this paper first for each module separate experiments,combined with overall system experiment,the experimental results compared with the test alone,although did not meet the demand of industry,but proved the validity of the designed system,has the potential for improvement.