Research On Automatic Reading Algorithm Of Pointer Meter Based On Deep Learning

Owing to their inherent stability and reliability,the pointer meters fin d extensive use in a diverse range of scenarios,including data recording and monitoring,rendering them an indispensable tool in the realm of production.However,the conventional method of manually recording readings from pointer meters poses issues such as exorbitant costs,low efficiency,and high-security risks,thereby failing to meet the requirements of modern product development and life.The realization of the automatic reading of pointer meters is in line with the development trend of intelligence in our country.This thesis addresses the situation where manual acquisition and labeling of datasets is costly and there are multiple types of pointer meters on the images acquired in real environments.The main contents are as follows:1.For multiple types of pointer meter detection tasks,a pointer meter data augmentation method for convenient generation of training datasets for deep learning networks is constructed,and a region-constrained strategy for solving the classification confusion problem in meter detection is proposed.When faced with a lack of dataset,the costly and inefficient manual acquisition and labeling method is often used.This thesis proposes a data augmentation method that can convenient generate a large number of labeled pointer meter data with only a few sample images.Based on this data augmentation approach,a set of pointer meter data set is produced which can be applied to the training of subsequent meter object detection network.In addition,since there is more classification confusion when using the YOLOv5 object detection algorithm for meter detection,the region constraints strategy in this thesis improves the classification accuracy in meter detection by adding significant regions on the dial of each class as auxiliary classes in the training to widen the variance between classes.2.To achieve a stable reading of meter dials intercepted by object detection,an algorithmic process for automatic reading of pointer meters based on feature match is proposed.Firstly,through the perspective transformation based on the AKAZE algorithm,the image of the dial to be read is aligned.Secondly,the dial image is unwrapped into a rectangle by utilizing the polar coordinate transformation method combined with the pre-obtained dial information.Then,the image is binarized to extract the pointer area,and the location of the pointer line is obtained by accumulating and comparing the pixel values.Finally,the distance method is used to calculate the number of indications to achieve automatic and stable readings of pointer meters.3.An improved YOLOv5 algorithm is proposed for the subsequent deployment application of the automatic pointer meter reading algorithm to achieve lightweight of the meter object detection network.Firstly,Shuffle Net v2 is selected as the backbone network to achieve the balance of speed and accuracy in feature extraction.Secondly,a GG-C3 structure is constructed to replace the C3 module in the network based on the G-Ghost module to reduce feature redundancy.Finally,the GSConv convolution is used to replace the standard convolution in the neck to further reduce the number of network parameters and computation while stabilizing the detection accuracy.The experiments show that the improved YOLOv5 algorithm proposed in this paper achieves the lightweight of the algorithm without affecting the overall automatic reading accuracy,so that the applicability of the automatic reading algorithm is improved.