positioning is based on the number and geometric features of the number box. However, such a detection method has obvious disadvantages. It is easily affected by
bubbles on the dial of the water meter, the angle of inclination of the image, blurring,
etc., and the detection accuracy is low.
The method based on projection is a method of combining the gray values of each
pixel of the obtained binarized image in a certain cross-sectional direction. The projection method is to count the number of black pixels in the horizontal and vertical
directions displayed in the image reading area, and draw a corresponding projection
image according to the counted number of pixels, and then locate the position where
the number on the dial is located. Specifically, the horizontal projection is to judge the
pixel value of each row of the obtained binarized image, and the number of gray values
of all the black pixels is counted; the vertical projection is to judge the pixel value of
each column of the obtained binarized image and count the number of gray values of all
the black pixels of the column. However, such detection methods are also susceptible
to many factors, such as light spots, stains, partial occlusions, etc., on the dial of the
water meter, and the accuracy is also low.
In this paper, the convolutional neural network is used to predict the reading area,
and then the non-maximum suppression algorithm (NMS) is used to remove highly
overlapping results from prediction region results to obtain the position of the reading
area. The IoU of the images in the 1000 test sets are all above 0.8 and combined with
the three-layer BP neural network for character recognition, the accuracy rate is 98.0%.
Compared with other methods, the reading results of the method in this paper are not
easily interfered by external factors such as bubbles, spots, blurs, etc., and the sensitivity is low.
The paper is organized as follows. Section 1 is talked about introduction to this
research. Section 2 introduces the specific implementation method of the predicted
reading area. Section 3 contains the experimental results and analysis. Section 4
summarizes the paper.
2 System Implementation
The network structure of the method is shown in Fig. 1. The implementation mainly
includes the following parts. Firstly, the convolutional neural network is used to predict
the reading area. Secondly, the non-maximum suppression algorithm (NMS) is used to
remove highly overlapping results from prediction region results. Finally, the accuracy
is calculated by the result of character recognition and thus compared with other
algorithms.
2.1 Predict the Reading Area with Convolutional Neural Network
To perform character recognition on the readings in the water meter image, the first
thing to do is to detect the correct reading area and eliminate character interference in
the non-reading area, which is critical for character recognition of the readings.
The specific structure of the convolutional neural network is shown in Fig. 1. It
consists of several convolutional layers and several pooling layers, including five two174
J. Wu et al.
bubbles on the dial of the water meter, the angle of inclination of the image, blurring,
etc., and the detection accuracy is low.
The method based on projection is a method of combining the gray values of each
pixel of the obtained binarized image in a certain cross-sectional direction. The projection method is to count the number of black pixels in the horizontal and vertical
directions displayed in the image reading area, and draw a corresponding projection
image according to the counted number of pixels, and then locate the position where
the number on the dial is located. Specifically, the horizontal projection is to judge the
pixel value of each row of the obtained binarized image, and the number of gray values
of all the black pixels is counted; the vertical projection is to judge the pixel value of
each column of the obtained binarized image and count the number of gray values of all
the black pixels of the column. However, such detection methods are also susceptible
to many factors, such as light spots, stains, partial occlusions, etc., on the dial of the
water meter, and the accuracy is also low.
In this paper, the convolutional neural network is used to predict the reading area,
and then the non-maximum suppression algorithm (NMS) is used to remove highly
overlapping results from prediction region results to obtain the position of the reading
area. The IoU of the images in the 1000 test sets are all above 0.8 and combined with
the three-layer BP neural network for character recognition, the accuracy rate is 98.0%.
Compared with other methods, the reading results of the method in this paper are not
easily interfered by external factors such as bubbles, spots, blurs, etc., and the sensitivity is low.
The paper is organized as follows. Section 1 is talked about introduction to this
research. Section 2 introduces the specific implementation method of the predicted
reading area. Section 3 contains the experimental results and analysis. Section 4
summarizes the paper.
2 System Implementation
The network structure of the method is shown in Fig. 1. The implementation mainly
includes the following parts. Firstly, the convolutional neural network is used to predict
the reading area. Secondly, the non-maximum suppression algorithm (NMS) is used to
remove highly overlapping results from prediction region results. Finally, the accuracy
is calculated by the result of character recognition and thus compared with other
algorithms.
2.1 Predict the Reading Area with Convolutional Neural Network
To perform character recognition on the readings in the water meter image, the first
thing to do is to detect the correct reading area and eliminate character interference in
the non-reading area, which is critical for character recognition of the readings.
The specific structure of the convolutional neural network is shown in Fig. 1. It
consists of several convolutional layers and several pooling layers, including five two174
J. Wu et al.