Research And Implementation Of Manufacturer Label Positioning And Text Recognition System On Industrial Trays Based On Deep Learning

Machine vision is a branch of the rapid development of artificial intelligence.It uses machines to simulate human vision to obtain useful information from images.It has a wide range of applications in industrial production and daily life.With the further development of modern intelligent manufacturing technology,higher requirements are put forward for the application scenarios of machine vision,traditional machine vision object detection and text recognition technologies usually use manual methods of extracting object image features,which are less robust,Unable to adapt to the diversity of different backgrounds and forms of illumination.To this end,this article combines the actual needs of the industrial automation field,that is,to detect the manufacturer’s label and its text on the industrial electronic tray,and conduct research on machine vision object detection and text recognition technology based on deep learning.The specific research and development work is as follows:(1)Positioning of industrial electronic trays,manufacturer labels and hollow areas.By consulting related papers on object detection,the detection accuracy of the Mask R-CNN model is higher than that of YOLO and Faster R-CNN,so Mask R-CNN model is selected as the basis of the object detection model.First,collect the data images required for training the model,and then use the VIA annotation tool for annotation.After designing the network model of this function,input the data,train the model and get the weight file.In the continuous iteration,the accuracy of the model is improved by data enhancement,parameter tuning optimization and other methods.In terms of model structure,Res Net50-FPN is used to replace the original backbone network Res Net101-FPN,and a set of new size anchor frames are added to detect manufacturer tags.Experimental results show that the improved model has a m AP value of 98.5% and a prediction speed of 7fps.(2)Self-adaptively find blank areas.After obtaining the industrial electronic tray,the manufacturer label on the tray and the coordinate information of the hollow area through the Mask R-CNN model,it is necessary to find a blank area of 10mm*10mm size on the industrial electronic tray to paste the new manufacturer label.The algorithm designed in this article is similar to a sliding window scanning the entire picture,and counting the pixel information in the window.If the sum of the array values in the window is 0,it means that the window does not overlap with other areas,indicating that the area is a blank area.Affix the new manufacturer label.(3)Text detection of manufacturer label.There are many important text information on the manufacturer’s label.This article uses the CTPN model for text detection.CTPN introduces a cyclic neural network to learn the sequence information of the text,and fully mine the feature information of the text to better locate the text position.This article considers that the height and width ratio of the character is roughly 2:1,so the 2*1 convolution kernel is used in the pooling layer to replace the 2*2 convolution kernel,which can better fit the shape characteristics of the characters and retain the character information in the image.The evaluation index m AP value of the improved model increased by 3.2%.(4)Text recognition of manufacturer label.This thesis uses the CRNN+CTC model for text recognition.By using the CTC loss function and introducing a blank character strategy,the automatic alignment of image input and text output is realized,and the complex character segmentation and other preprocessing operations are simplified,so as to realize the end-to-end recognition of the model.At the same time,this paper uses a GRU network with fewer parameters to shorten the training time of the model.The accuracy of text recognition reached 98.3%.