Research And Application Of Water Meter Reading Recognition Method Based On Deep Learning

Unlike intelligent water meters such as the Internet of Things(Io T)water meters,wireless remote transmission cold water meters,and ultrasonic water meters,mechanical water meters require manual meter reading,which is boring,inefficient,easily affected by the physical and mental state of the meter reader,and sometimes even affected by interpersonal relationship.An effective solution is to detect and recognize water meter readings by using computer vision and image processing technology.However,images of water meter dials captured in natural scenes might have problems such as different lighting,blurred by dust,tilt,fog,and non-standard pictures.These problems bring challenge to the detection and recognition of water meter readings.In recent years,rapid development of deep learning,especially target detection algorithms,has made it possible to detect and recognize images in natural scenes.In such a context,this paper focuses on the deep learning-based detection and recognition of mechanical water meter readings.This research mainly does the following work:1.Considering limited public data sets of water meters in complex scenarios,a data set of water meter dials in natural scenarios is established.The data set established herein includes 1,563 images of various complex scenes(under different lighting,blurred by dust,tilt,fog,shadow and other conditions).Samples in the data set are then enhanced to further enrich the water meter data set.2.Addressing such problems with images of mechanical water meters in natural scenes as complex environment and low detection accuracy,this paper introduces a YOLOv4 optimization-based target detection algorithm.First of all,Focus structure is introduced to the feature extraction network of YOLOv4 to form a new feature extraction network,improving its receptive field while reducing the loss of original information.This also solves the problem of low resolution and difficult recognition of water meter readings.In view of the unsatisfactory lighting conditions in nature and dirt on the surface of the water meters,the structure of the feature fusion network is redesigned to improve the feature expression and reduce the computational complexity.In the end,the loss function EIOU is introduced to further enhance the detection accuracy.The improved network model reaches the detection accuracy of 95% and m AP as high as 98.1%.The parameter amount and calculation amount of the proposed model are reduced by 48.6% and 36.8%.3.To overcome the difficulties in mechanical meter reading and reduce the cost thereof,a meter reading management system is designed to achieve intelligent meter reading management.Firstly,a demand analysis is given to the water meter identification system so as to facilitate the design of the overall architecture and application system.In the meantime,the optimized model is tested through Open VINO inference framework.According to the results,its detection speed meets the application requirements of real-time detection.Next,deep learning and relevant computer technology help achieve(i)uploading water meter pictures and readings via mobile phone APP,checking the readings on the server-side with algorithms called,(iii)displaying front-end services and APP service module interface.The corresponding functions are also expounded,demonstrating the level of intelligence of the meter reading process.