Research And Implementation Of Door Check Detection And Recognition Based On Machine Vision

With the continuous development of machine vision technology,there are more and more industrial application in this field.In the process of door production,after the door check is installed,it is necessary to confirm that the installation is correct.Nowadays,the installation of the part is judged by human eye observation,and the continuous work inevitably causes the decline of the correct rate.In order to solve this problem,it carries out related research on the technology of machine vision in the detection of door check.The algorithm framework is divided into four parts:(1)For the screw nut on the part,the Adaboost frame is used to accurately position the nut.The LBP feature of the nut is used to train multiple binary tree weak classifiers,and all weak classifiers are combined into a strong classifier.The strong classifier is used to locate and identify the nut.(2)Positioning and segmenting the characters on the part.The characters on the door check are the concave characters.It is difficult to accurately locate them by existing method.Therefore,for this kind of character,a strong light source is used to forward illuminate the characters to highlight the edges of the characters,and then the improved method is used to locate and segment the characters.Before the positioning and segmentation,the improved Canny algorithm based on prior knowledge is used to extract the character edges.When positioning and segmentation,two methods are used: edge-based character localization and segmentation,MSER-based character localization and segmentation.The first algorithm accurately locates characters according to the edge density of characters,and then uses vertical projection and connected domain methods to segment the characters.The second algorithm uses MSER to locate characters on the edge enhancement map of the part,and then uses the connected domain method to segment the characters.Experiments show that the edge-based character localization and segmentation method is superior to the MSER-based character localization and segmentation method in accuracy,but it is not as fast as the MSER-based character localization and segmentation method.(3)Identify the characters on the part.It uses is by three methods to identify the characters: character recognition based on multi-template matching,single character recognition based on convolutional neural network,and end-to-end character recognition based on Faster-RCNN.The first method creates a template library by supervised learning firstly.And then accelerate template matching by using an efficient algorithm based on BPC idea secondly.The second method builds a seven-layer simple network model: ‘convolution-pooling-convolution-dropout-pooling-convolution-full connection’.It uses the first six layer to extract feature and uses the last layer to classify the characters.The third method skips the character positioning phase and implements end-to-end character recognition.Experiments show that the recognition accuracy of the first method is the lowest and the recognition accuracy of the third method is the highest of these methods.