Research And Implementation Of Real-time Identification Method Of Digital Display Instrument

In recent years,the intelligence level in the industrial field has been continuously improved,and the industrial upgrading has greatly improved the production efficiency.As a widely used industrial accessory,the intelligence of instrument production testing has also received unprecedented attention.Intelligent instrument test refers to the use of computer-related technology to achieve the inspection and test of instrument functions,so as to ensure the product quality.Product quality is the foundation of the industry,and the last level to ensure product quality is product testing.However,traditional manual testing instruments have the disadvantages of high cost and unstable accuracy.How to efficiently and realize instrument testing accurately has become a problem that many instrument manufacturers think about.Therefore,the researchers are committed to using image recognition knowledge,give the task of judging whether the normal display interface of the instrument to the computer to complete,and achieve the effect of real-time detection,and fundamentally optimize the cost,efficiency and quality of the instrument production,so as to improve the core competitiveness of the industry.The instrument is a target with stable morphology but complex interface.Traditional methods mostly use target detection methods to extract the instrument information and then conduct identification analysis.However,the current methods have encountered difficulties in the real-time and robustness of the identification,and they are not mature enough in practical application.For the above problems,this paper proposes a real-time instrument test method based on image recognition,and develops a realized instrument test system.The present method is divided into two parts: target detection and information identification.First,the improved YOLOv3 algorithm is positioned to extract the instrument area,then uses the relevant method of digital image processing,and uses the projection method to extract the instrument panel and instrument information successively according to the edge features.Finally,we identify the digital pipe,indicator lamp and instrument model respectively through the improved wiring method,area method and OCR technology.In the instrument test process,the following improvements are proposed in the paper.(1)According to the complex instrument detection background and fast speed,a target detection algorithm based on improved YOLOv3 is proposed.Based on the YOLOv3 algorithm,two Dense Net(Densely Connected Convolutional Networks)network blocks were first added to the Darknet to enhance the reuse of features by the network.Then,Darknet was used as a feature extraction network,performed model pruning,adjusted the number of residual networks to 4,modified the partial convolutional neural network to a deep separable convolutional network,and then modified the convolutional 6 layers before all detection layers(YOLO Detection)to 2layers to reduce the parameters of the model.At the same time,the GDIOU(generalized-IOU and distance-IOU,GDIOU)boundary frame is proposed to return to the coordinate loss,and to readjust the weight of the loss function according to the detection requirements.Experimental results show that the number of improved YOLOv3 algorithm parameters is reduced by 30% compared to the original algorithm,the average accuracy in instrument detection by 2.75% and the detection speed by22.91%,which is beneficial to improve the real-time and robustness of instrument testing.(2)In order to improve the adaptation degree of the image threshold segmentation,an adaptive binary method based on the bimodal method and the Otsu method is designed.First,The gray-scale distribution histogram of the images was counted,and then the two best peaks were selected based on the two-peak method,while not calculating the gray range of interclass variance with 0-255 as the OSTU method,but directly entering the gray scale corresponding to the two peaks as the boundary.The proposed scheme can remove some unnecessary grayscale pixels in advance,and can also reduce the computation of the OSTU method while improving the accuracy.(3)In response to the current situation of low accuracy and poor robustness of identifying digital tubes,a modified sewing method based on boundary features is proposed.The state of digital tubes is directly judged by counting the number of boundaries.Compared with the original method of pixel number statistics,the boundary features are relatively stable,which can improve the stability of the method.The improved method improves the recognition speed by about 20% and improves the accuracy by 2.1%.In general,this paper takes digital instrument image processing based on experiment,instrument function test as the research goal,to achieve high precision,real-time as the research orientation,through the design and development of a complete instrument test system,experiment and application in self-made data set,to achieve a complete real-time instrument test task.