Research on Automatic Reading Recognition of Wheel Mechanical

Water Meter Based on Improved U-Net and VGG16

LIUKUI CHEN1, WEIYE SUN1, *, LI TANG1, HAIYANG JIANG1, ZUOJIN LI2

Chongqing, 401331, CHINA

Abstract: This paper proposes a deep learning scheme to automatically carry out reading recognition in wheel

recognition accuracy of 95.6% is achieved on the dataset in this paper. This paper proposes a new solution for

the current situation of manual meter reading, which is time-consuming and labor-intensive, errors occur

frequently; and the transformation cost is high and difficult to implement. It provides technical support for

automatic reading recognition of wheel mechanical water meters.

Key-Words: Wheel Mechanical Water Meter, Reading Recognition, U-Net, VGG16

Received: September 23, 2021. Revised: June 21, 2022. Accepted: August 13, 2022. Published: September 1, 2022.

public facilities built before 2000 in Chongqing.

Chongqing Tap Water Company owns nearly 2

million users of water, and mechanical water meters

account for nearly 60% of the installations of these

users [5]. The "one household one meter" policy

implemented in succession throughout the country

represents the increasing attention of the state and

society to saving water resources, but at the same

time, it greatly highlights the cumbersome and

inconvenient nature of the traditional meter reading

reading method are becoming increasingly

prominent. It is not only time-consuming and

laborious, this also happens frequently that meter

readers cannot enter narrow and rugged areas, which

makes it difficult to collect data. Moreover, due to

door to door is very heavy [12]. At present, there are

also ways to transmit data and power at the same

time based on existing wires. For example, Ben-

shimol introduced an effective method of automatic

meter reading from smart meters by using a power

line communication network, and two intelligent

polling algorithms based on application layer

methods are used to ensure the effective

transmission of data in the network [13]. This

method has a simple structure, mature technology,

purchasing a certain amount of water card in

advance [14]. This method is simple to use and

convenient to replace and install. Its main

shortcomings are: on the one hand, due to the lack

of information transmission devices to connect users

and companies, timely water supply statistics and

dispatching will be difficult to achieve [2]; On the

other hand, the economic losses of users or water

supply companies caused by the failure and damage

of water cards or malicious modification of users'

improve and innovate the traditional methods has

gradually become the research focus of relevant

scholars at home and abroad.

At present, the automatic recognition technology

of water meter reading based on image processing

has been theoretically studied, but it has not been

applied on a large scale in the market. For example,

Jing-wei Sun combined the color characteristics of

the pointer, used the global threshold and local

threshold to segment the water meter components,

obtain the segmented single character matching

result [16]. Tian-hua Liu transformed the water

meter pictures into HSV color model, extracted the

H-channel, removed the noise, and obtained the

contour by using median filter and canny operator,

and then calculated the center coordinates of the

character, and then carry out numeral recognition.

Firstly, the water company outsources patrol

personnel of the old residential compounds to

collect the reading images of water meters in a

positioning algorithm, the reading area in the dial

image is extracted and the irrelevant area is

removed. In the next place, segment each character

in the target area, then recognize the obtained single

character to get the reading result. After that, the

results will be sent to the billing center. Finally, the

water bill will be returned to each user after being

summarized and sorted by the billing center. The

basic logic block diagram of this remote meter

reading system is shown in Fig.1.

Where

is identity mapping, that is, the output

of the previous layer of the network,

is

residual mapping, that is, the output of the current

layer of the network, and

is the network

mapping from the input to the sum. Adding identity

operations in the encoder and decoder are included

in the residual blocks to perform feature extraction

and step-by-step up-sampling of the target area of

the dial respectively. The shallow convolution

focuses on the texture features of printed characters

in the reading area, and the deep convolution

focuses on the essential features. In the meantime,

multiscale feature fusion is performed between the

encoder and decoder, such fusion connections run

through the whole network, so that the final up-

preprocess and expand the samples of water meter

dataset by means of filtering and rotation

transformation, so as to enhance the generalization

prevent overfitting problems in the model, and

greatly reduce the memory occupation of VGG16

model [29].

The last convolution layer of the designed ten

classification VGG16 network shall output 10

feature maps. Using the global average pooling

instead of the full connection layer to directly

associate the feature maps of the last convolution

layer with the categories, respectively accumulate

all pixel values of each feature map and calculate

makes the network more robust to the spatial

transformation of the input image.

Fig.5 Replacing the full connection layer (up) with

GAP (down) in VGG16

set to the number corresponding to the character;

channels of each pixel of the image according to a

certain weight to obtain the average value, the

calculation formula is shown in formula (2). Where

represents the gray value result of any

point of the image based on the weighted average

method,

,

, and

are the

Table 1 Factors affecting reading segmentation and

Fig.7 Pretreatment images of water meter (left) and

non-target area on the right side. After restoring the

shape and size of the rectangular reading box area,

the target region extraction result is obtained by

performing the "and" operation between this image

and the original image. As shown in Fig.8:

area extraction result image (right)

4.3 Reading identification

After obtaining the extraction result of the reading

and combined with the research object of this paper,

there is an obvious black vertical line connecting the

top and bottom between each reading character of

the mechanical water meter. Firstly, this paper

detects the tilt angle of these parallel interval

vertical lines through Hough transform to correct

the tilt of the reading area. In the standard

parameterization mode, the straight line can be

expressed as

distance from the straight line to the origin, and

is

the angle between the straight line and the x-axis.

The straight line detection of Hough transform is

realized by determining the intersection point of the

curve transformed by the point on the straight line in

with the same spacing distance, the reading string

can be divided into single characters. The results are

shown in Fig.9:

character segmentation diagram (right)

Fig.10 The identification result after splicing

Table 2 Comparison of parameters between

Fig.12.

Fig.11 Statistics of the number of reading characters

In the process of reading area extraction

experiment, this paper also uses YOLOv5 target

detection algorithm to compare with the improved

of detecting read regions using the YOLOv5

algorithm, where the pre-trained model

YOLOv5s.pt is used, and the number of epochs for

training is 150.

segmentation and positioning of U-Net, and

algorithms such as line detection still need to be

carried out to extract exact edges of the reading

area. At the same time, YOLO algorithm has higher

requirements on dataset, model volume and

complexity than U-Net [33]. The water meter

dataset with small samples has high similarity, and

the background similarity is also high, the type to be

identified is single. Thus, it is more accurate and

applicable to directly use U-Net, which is simple,

weak.

In the algorithm of this paper, the character

region is accurately located and segmented to

remove the influence of the background region and

the separation line; Then combined with the "half

character" processing scheme, the single character

recognition is carried out, which solves the reading

recognition problems in special cases such as the

distortion of the number, the size difference and the

deformity caused by the rotation of the gear, and

4.5 Discussion

In this study, the recognition strategy from coarse to

fine is adopted. First, the dial reading area is

accurately located at pixel level through the

improved U-Net network, then the single character

segmentation is carried out according to the

structural characteristics of the reading area, and

then the character recognition is carried out through

the improved VGG16 network. The experimental

results show that the technical scheme is very

and do not include the pointer type water meter;

2. It is difficult to extract effective features by

image processing in the case of too low brightness

and serious obscuration of sundries;

3. The structure of the designed segmentation

and recognition networks is still complex, and there

are still some limitations in real-time processing,

there is still room for improvement in accuracy.

The future development in the field of instrument

identification requires higher resolution and picture

recognition of the pointer type water meter will be