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

Meter reading is a heavy and important task for water operations department.The traditional manual meter reading method is time-consuming and laborious,and it is too expensive to switch to the digital measuring meter.Point water meter reading recognition based on computer vision technology is an effective alternative technology without replacing the original water meter.The traditional image recognition method is inaccurate and slowness.Meanwhile,the research on detection and recognition method of pointer water meter reading based on deep learning attracts people’s attention.This thesis aims to research on detection and recognition method of pointer water meter reading based on deep learning.The main works are presented as follows:1.Build a pointer water meter database covering practical application condition.It includes:developed a set of acquisition criteria for pointer water meters in real scenes which could be used for the training of detection and recognition model;proposed a set of standardized processing method with less collection workload for a few non-standard pointer water meter images,which including rotation,cutting and other main steps;designed a labeling method for the pointer water meter image based on the pointer direction interval judgment to facilitate the detection and recognition.2.Proposed a method to remove dust and mist from water meter images based on dark channel prior and DehazeNet.This method designed a computational method for the atmospheric light components in the water meter scene using the scene information and water meter type information;it improved the computational method of the water surface scene transmittance for the characteristics of the water meter scene;it used dark channel prior calculation process to remove dust and mist.The experimental results show that the mean average precision(mAP)was improved by 0.012 and the error rate was reduced by 12.77%by using this model,which verifies the effectiveness of the method.3.In order to realize the detection and recognition of the pointer water meter image,this thesis proposed a detection and recognition method based on R-FCN,and constructed a water meter reading recognition system.This method included feature extraction module,RPN module,multi-scale feature detection module,classification module,and position-sensitive ROI pooling module.The RPN module added a residual block structure to ensure the more accuracy of the extraction of the region of interest.A multi-scale feature detection module was added between the feature extraction module and the classification module,and simplified the difficulty of subsequent detection and recognition by learning the different scale features.The classification module combined the results of the multi-scale feature detection module and an added residual block structure,improved the accuracy of the detection and classification of the water meter pointer effectively and reduced the accuracy degradation caused by the network layer depth.The mean average precision of the proposed detection and recognition model by this thesis is 0.906,which is 0.024 higher than the original R-FCN model,and the error rate is reduced by 20.34%.Meanwhile,the mean average precision could be 0.918 by using the proposed model which can remove dust and mist from water meter images.The pointer water meter detection and recognition method proposed in this thesis realizes the accurate detection and recognition of the pointer It has obvious advantages in performance and speed comparing with other deep learning network and the traditional detection and recognition method.