Pointer Dial Reading Recognition Method Based On Deep Learning

With the rapid development of deep learning,all walks of life began to accelerate the use of machines instead of manual operations.As a necessary measurement tool for life,the instrument will be inefficient and the error rate will increase due to weather factors,the existing environment and the uncertainties existing in manual meter reading,so its reading operation is gradually beginning to change from manual transcription to machine execution.At this stage,the proposed dial recognition algorithm can only be read for a single dial style under specific conditions,and there are low detection efficiency and poor accuracy.In view of the above problems,this paper will study the method of drawing the reading of the dial from the aspects of image preprocessing,dial area extraction,dial scale and pointer segmentation,and obtaining readings.The main contributions of this paper are as follows:(1)Extraction method of analog instrument dial.In the image preprocessing part,this paper uses histogram equalization to correct the phenomenon that the image is too bright or too dark,adaptive median filtering to remove noise,perspective distortion to correct some special pictures collected by fisheye lenses,improve the accuracy of subsequent model image processing,and improve the model effect.Based on the object detection algorithm of YOLOv5 s,a new feature extraction network is proposed,which replaces the C3 module by using the Ghostdial module in the backbone feature extraction network,which increases the recognition speed of the network without affecting the network detection accuracy.This module uses a small amount of computation through traditional convolution to generate a feature map with a small number of channels,and then generates a new feature map through deep convolution based on the feature map,and finally stitches the two feature maps to obtain the output of the final picture.Experiments show that the improved YOLOv5 s object detection model improves the inference time by 4ms/frame based on satisfying the detection accuracy.(2)Analog dial reading recognition method for a specific measuring range.For the dial data of a specific range,this paper improves based on Deep Lab V3+ to extract the dial scale and pointer area,and improves the detection accuracy of the model while reducing the number of parameters by replacing the backbone network with a Res Ne St network,and increases the channel attention mechanism at the decoder end to increase the accuracy of network recognition accuracy)is up 3.25% from the original algorithm and 1.62% from MIo U.The dial scale and pointer area extracted by the model are expanded into rectangles by Cartesian and polar coordinate conversion,and the dial reading extraction operation is carried out by using the distance method to complete the pointer dial reading extraction for a specific range.(3)Adaptive reading recognition method for analog dials.This paper proposes an adaptive dial reading extraction method to adapt to various measuring range analog dials.For the dial data,this paper uses Optical Character Recognition(OCR)to extract the dial digital range based on extracting the dial scale and pointer area,and then uses the adaptive dial reading extraction method to extract the reading.The adaptive dial reading extraction methods are mainly as follows: Correlate the detected text range area with the dial length scale;Divide the long scale marked with the range left and right according to the position of the pointer and pair it in pairs to calculate the interval range separately;The paired scale and the pointer calculate the angle separately according to the angle method,and calculate the reading according to the angle method;Several sets of readings are obtained according to the associated range,and the median of the reading result is used as the reading result of the dial.Experiments show that the OCR pre-trained network model based on Baidu framework can well meet the accuracy and speed of dial text extraction,and the adaptive dial reading extraction method can effectively control the error between the real reading,and the error result can be reduced to ±5% of the real reading89.30% in the test set,and the recognition accuracy can meet the needs.