An intelligent method of energy monitoring and analysis based on image recognition technology (2023)

With the rapid development of society and the growing demand for energy, the importance of energy monitoring and management is becoming more and more visible [1-5]. Smart energy monitoring and analysis based on image recognition technology can provide more accurate real-time data support for the energy system and improve the efficiency and level of energy management [6-13]. Against the background of Industry 4.0 and digital transformation, all segments of life are constantly evolving towards intelligence. The energy industry also needs to move with the times and use technical means such as image recognition to intelligently monitor and analyze energy. In all combinations of energy production, transport and use, it is necessary to monitor the operation of devices and safety hazards in real time [14-21]. The use of image recognition technology can quickly and accurately identify abnormal states of power equipment, thereby improving the work efficiency and safety of power equipment. It is also necessary to study this method meaningfully.

Imran et al. [22] evaluated and compared the state-of-the-art You Only Look Once (YOLO) deep learning algorithm with traditional manual text extraction and recognition functions. The image dataset contains 10,000 electricity meter images, further augmented by in-plane rotation and upward data to make the deep learning algorithm resilient to these image variations. For training and evaluation purposes, the image dataset is flagged to create a baseline truth for all images. The proposed method can be very helpful in reducing the time and effort currently associated with meter reading, i.e. staff visiting homes, photographing meters and manually extracting readings from these images. To solve the problems of low efficiency and low quality of manual sorting in traditional meter recycling, Hu et al. [23] used digital image processing technology to build an automatic sorting system for meter recycling based on an artificial neural network. First, the basic requirements of the system construction were analyzed in detail, and then the principle and method of image recognition using an artificial neural network was presented in detail. On this basis, the entire skeleton of the automatic sorting of the recycling of electric meters was built. Finally, the design and application of functional modules were made and the Azure database was built via the SQL Server platform, which implemented the system application of this research. The final application shows that the automatic sorting system built according to the research results is characterized by a simple interface and easy operation, which can significantly improve the efficiency and quality of meter recycling and sorting, and has some practical significance for the development of the State Grid. industry. Bajaj and Yemula [24] proposed the Instrumental Image Capture and Processing System (MICAPS). MICAPS is an image-based data extraction technique. MICAPS is implemented by integrating the Raspberry Pi with the Raspberry Pi camera. Python code is used to configure the camera, take photos and send them to the cloud. This image is a seven-segment display. Optical computer image recognition algorithms are used to convert a seven-segment display into text. Kanagarathinam and Sekar [25] describe collecting sets of raw images of digital electricity meter screens, detecting text and recognizing seven-digit segments from collected samples, which can help to reduce the costs of Advanced Metering Infrastructure (AMI). The proposed dataset has great potential for electricity billing based on fully automatic optical character recognition (OCR).

At present, the image data recognition methods of energy monitoring instruments for intelligent energy management and analysis mainly include: template matching method, edge detection method, feature extraction method and deep learning based methods. Most of these methods are sensitive to factors such as input image quality, lighting, angle, etc., and recognition performance may be poor when image quality is not high. Some methods, such as feature extraction and deep learning, require a large amount of computation and relatively poor real-time performance, which can affect the timeliness of energy monitoring. In addition, existing methods work well for some types of energy monitoring instruments, but may have poor generalizability for new or other instrument types that require re-training or re-tuning of the model. For this purpose, the study conducted research on intelligent energy management and analysis methods based on image recognition technology. Part 2 of the study involves pre-processing the panels with energy monitors using brightness control and Hough transform. After removing the dial of the pointer instrument in section 3, the combined domain analysis, thinning algorithm, line fitting and pointer direction estimation mechanism are used to detect the pointer and calculate the angle. The fourth part shows how to recognize the readings of the electricity monitoring device. The effectiveness of the proposed method was confirmed by the analysis of experimental results.

picture 1.General architecture of an intelligent energy management and analysis system

In the actual intelligent power management analysis scene shown in Figure 1, due to the complexity of the recording environment, the light easily affects the recording process of the analog meter, which causes great interference in the automatic identification and reading of the meter. . instruments later. To improve the accuracy of automatic identification of instrument readings, image preprocessing is essential. In real-life scenes, it is often difficult to maintain constant lighting conditions, and adjusting the brightness helps to eliminate the effects of changes in lighting conditions and bring the image closer to the ideal state. Moreover, the image contrast can be improved to make the difference between the instrument panel, pointer and scale more obvious, which is convenient for later identification and positioning. The Hough transform is a method for detecting specific geometric shapes in images, such as lines and circles. With the Hough transform, the edge of the instrument's target can be efficiently detected and thus accurately locate the target. At the same time, a straight line in the image can be detected, which is useful for pointer positioning and detection, and then serves as the basis for pointer angle calculation and reading calculation.

Histogram equalization(THAT) IContrast Limited adaptive histogram equalization(conflict) can be used to adjust the brightness of the Energy Monitor image. However, in the real scenario of intelligent energy management and analysis, the useconflictAdjusts the brightness factor of the power consumption monitor imageTHATsuch as local brightness adjustment, preventing excessive noise, preserving image detail information, avoiding local saturation, etc. These advantages help improve the image quality and further improve the energy monitoring instrument's image data recognition accuracy and performance. The basic flow of the algorithm is as follows:

(1) Convert the input color image to luminance-chrominance space so that only the luminance component can be processed in subsequent steps without affecting the color information.

(2) Split Areas: Divide the brightness component image into multiple small areas (tiles).

Local histogram equalization: equalizes the brightness histogram of each small area. This step can improve the contrast of local regions while preserving detailed image information.

(3) For the brightness histogram of each small region, each gray level corresponds to a set. If the set size is greater than the specified threshold, grayscale pixels that exceed the threshold are removed. Suppose the gray level of each processed small area is expressed asI_G, the width of the small image is expressed asask_A, the height of each small surface is expressed asask_B, then the average pixel can be calculated using the following formula:

$X l Q=\frac{Q_a \times Q_b}{I_g}$ (1)

The actual threshold can be calculated using the following formula:

$V_g=I_{p j} \times X l Q$ (2)

Removed grayscale pixels are evenly spaced. Suppose the interval at which pixels are removed is expressed asV, the total number of pixels removed is expressed asPotassium, the maximum gray level of the image is expressed asV_C, then the sliding interval can be selected where:

$x_l=\frac{K}{I_g}$ (3)

$V=\frac{V_c}{K}$ (4)

(4) Contrast Threshold: To avoid excessive noise amplification, the contrast threshold can be set. When performing local histogram equalization, gray levels above the threshold are clipped, and the trimmed gray levels are evenly distributed to other gray levels to reach the contrast boundary.

(5) Region fusion: Because each small region has undergone independent histogram alignment processing, unnatural edge transitions may appear. Therefore, it is necessary to perform a weighted average for the adjacent regions to achieve a smooth transition effect.

Specify the center point of each small area as a pattern point. allowo=C-C_-/C₊-C_-,vas=G-G_-/G₊-G_-.Suppose the gray pixel value (C,G) Expressed asC(N), the gray value of the four adjacent points around a pixel point is expressed asC_--,C_+-,C_-+, IC₊₊. Going through the image, the following gives the grayscale linear interpolation formula:

$\begin{wyrównane} & C(n)=o\left[u C_{--}(n)+(1-u) C_{+-}(n)\right] \\ & +(1+o )\left[u C_{--}(n)+(1-u) C_{+-}(n)\right]\end{对齐}$ (5)

(6) Convert back to original color space: combined luminance componentconflictManipulate the color components of the original image, then convert the image back to its original color space (e.gRGBspace).

After performing the above steps, after adjusting the brightness, you can get an image of the energy meter, which is used for subsequent image recognition and analysis.

photo 2.Schematic of Hough's circuit detection principle

In the real-world scenario of smart energy monitoring and analysis, the Hough transform is an effective geometric shape detection method for detecting a circular dashboard in images of energy monitoring instruments. For circle detection, the parameter space includes the coordinates of the center (A,B) sum radiusRThe Hough transform finds the combination of parameters most likely to represent the circuit by discretizing and accumulating in the parameter space. Before performing the Hough transform, you must first perform image edge detection and extract the dash board edge information. Then, according to the equation of the circle, draw a circle for each edge point of the image in parameter space, taking that point as a point on the circumference, and the center and radius as parameters. In the parameter space, circles corresponding to all edge points are accumulated and the combination of parameters with the highest cumulative value is found, i.e. the detected circular target. Figure 2 is a schematic detection diagram of Hough's circuit. Circle Detection Hough locates circles using a voting mechanism. In particular, the standard circuit equation (C-o)²+(G-vas)²=F²converted intotelcoordinate system, namely:

$o=c-f \cos \varphi ; u=g-f \sin \varphi$ (6)

Many energy monitoring instruments use pointer readings, and the position and angle of the pointer directly determine the accuracy of the measurement. By detecting the pointer and calculating its angle, accurate data can be obtained from the image, providing a reliable basis for intelligent energy monitoring and analysis. Traditional data collection from energy monitoring meters usually relies on manual reading, which is inefficient and error-prone. By detecting the pointer and calculating the angle, the automation and intelligence of data collection by energy monitoring instruments can be realized, improving the efficiency of data collection and reducing the number of errors. Existing methods for detecting the pointer and calculating the pointer angle have their advantages and disadvantages and can be affected by factors such as noise, lighting, and shape changes, leading to a reduction in detection and calculation accuracy. Therefore, the existing methods can be considered improved in practical applications.

For reliable recognition of image data from energy monitoring instruments in complex environments, this study used linked domain analysis, a thinning algorithm, a line fitting and a pointer direction evaluation mechanism to detect the pointer and calculate the angle after pointer extraction. Linked domain analysis can effectively detect areas of indicators with similar color and grayscale characteristics, adapting to indicators of different shapes and sizes. The thinning algorithm can extract the skeleton of the indicator and better describe the geometric features of the indicator. By aligning the lines, the angle of the pointer can be calculated more accurately, thus improving the accuracy of data recognition. Combined with the pointer direction estimation mechanism, the pointer direction can be automatically identified, realizing the automation and intelligence of energy monitoring instrument data collection. It can effectively reduce the errors caused by manual operations and improve the efficiency of energy monitoring and analysis.

This study breaks down pointer detection and pointer angle calculation into three steps: locating the pointer area and center of the instrument, determining the pointer-to-line alignment, and a mechanism for estimating the direction of the pointer. In the first step of locating the pointer area and gauge center, the goal is to determine the approximate position of the pointer in the gauge and the center of the dial. First, the image is pre-processed using image processing technology to extract the edge of the dial. Then use related domain analysis or contour detection methods to find pick areas. The center position of the instrument is then obtained by calculating the geometric center of the dial area or by finding the center using the circular Hough transform. Finally, by analyzing the color, grayscale, or edge information in the dial area, the pointer area can be predetermined.

The second step improves the alignment of the indicator and the line, with the aim of extracting the skeleton of the indicator and calculating its adjusted line. First, Zhang Suen's algorithm can be used to process the indicator area to get the indicator skeleton. This step removes the difference in the width of the pointer, making it one pixel wide. Then, the skeleton of the indicator is fitted by least squares to produce a line that represents the direction of the indicator. This line helps to calculate the angle between the pointer and the center of the instrument.

The Zhang-Suen algorithm is a classic algorithm for reducing the number of binary images. Extracts the skeleton of an image by iteratively removing edge pixels. When using the Zhang-Suen algorithm to extract the skeleton of the indicator region, it is important to perform some preprocessing on the image before running the algorithm to improve its performance. This may include noise removal, smoothing and sharpening procedures. The algorithm can generate imperfect skeletons such as small fractures and bifurcations. In practical applications, the results of the armature can be optimized by post-processing operations (such as morphological operations), and the accuracy of pointer detection and angle calculation can be improved.

When converting a cursor area image to a binary image, you must select the appropriate binarization method. Different binarization methods can lead to different matching effects, so it is necessary to choose the appropriate method according to the actual situation. To avoid infinite loops, the maximum number of iterations can be set as the termination condition of the algorithm. When the number of iterations reaches its maximum value, the iterative process should be stopped even if the image continues to change.

In addition, the skeleton of the indicator uses the method of least squares to fit the line. The method of least squares consists in subtracting the actual measurement value of the indicator skeleton from the theoretical value, and then taking the sum of the results of the least squares calculation. ifIadjustment point ((C_N,G_N), (N=1, 2...I), the actual point is expressed asvas_N, the expression for the matched line is:

$g(c)=o c+u$ (7)

The formula below gives the formula for calculating the sum of squared differences:

$h^2=\sum_{n=1}^i\left(g_n-g\left(c_n\right)\right)^2=\sum_{n=1}^i\left(g_n-\left( o c_n+u\右)\右)^2$ (8)

found valueoIvasprovide the minimumH², partial derivativeH²about the unknownoIvaswe find the partial derivative set to 0. The following formula gives the calculationoIvas:

$o=\frac{i \sum_{n=1}^i c_n g_n-\sum_{n=1}^i c_n \sum_{n=1}^i g_n}{i \sum_{n=1}^ i c_n^2-\lijevo(\sum_{n=1}^i c_n\desno)^2}$ (9)

$u=\frac{\sum_{n=1}^i c_n^2 \sum_{n=1}^i g_n-\sum_{n=1}^i c_n \sum_{n=1}^i c_n g_n }{i \sum_{n=1}^i c_n^2-\left(\sum_{n=1}^i c_n\right)^2}$ (10)

The third step is the pointer direction evaluation mechanism, the purpose is to evaluate the actual direction of the pointer and calculate the angle. According to the adjustment line obtained in the previous step, the angle between the line and the center of the instrument can be calculated. However, since the pointer has two ends, it is necessary to further specify which end is the actual end of the pointer. This can be done based on the length of the indicator, color, shape and other characteristics. For example, the tip of the indicator is usually shorter and darker in color. The exact angle of the pointer can be obtained using the pointer direction estimation mechanism. Table 1 shows the indicator direction evaluation mechanism.

Table 1.Indicator direction evaluation mechanism

Figure 3 shows the process of recognizing electricity meter readings. After the electricity meter dial is finished preprocessing and the electricity meter pointer is detected, it is necessary to remove the scale of the electricity meter. When a circular gauge face is detected, the edge features of the face are extracted using either the edge detection algorithm or the Sobel operator. In the case of a target edge image, the marks are detected using the Hough transform. When a scale line is detected, it is filtered and classified by characteristics such as length and angle. For filtered characters, calculate their angle relative to the center of the gauge. It should be noted that the angle calculation method must be adapted to the actual situation in order to adapt to different types of instruments. Depending on the sign angle and gauge range, the angle can be mapped to a real value. After calculating the pointer angle, match it to the extracted scale value.

photo 3.The process of identifying indications of electricity meters

Since the scale value usually appears on the large scale line, in this document the target image is segmented according to the detected location of the large scale line to obtain a local area containing the scale value. That is, set the search area near the main line of the scale to get the area where the scale value is located. The segmented local area requires some image preprocessing to improve recognition accuracy, including grayscale, binarization, noise removal, and smoothing. In the preprocessed image, the text area is in . Optical Character Recognition (OCR) technology is used to recognize a localized text area. The recognized tag values ​​are compared to the corresponding reference tags. This can be done by comparing the tick value area and the location of the major ticks. After the calibration is completed, the scale values ​​corresponding to the main scale lines on the meter are obtained.

In this article, after extracting the scale value image, the roller-based fast character recognition algorithm is further used to recognize the scale value. The wavelet transform can effectively capture the local features of the image, which helps to improve the accuracy of character recognition. Appropriate basic valence functions (such asThen,mixed raceetc.) and distribution levels can be selected to meet identification needs. The conversion of two-dimensional image data into one-dimensional data is achieved by vertical and horizontal mapping of the binary network character image. Projection data is decomposed into two wavelet layers, and feature information is extracted from the smooth components of each layer. Compare and analyze with the characteristic information of character templates for fast and efficient character recognition.

Suppose the valence function is expressed asOh_Other,Rice(C), the wavelet coefficients are expressed asG_Other,Rice.Expression of any functionR(C) in the spaceV²(F) is given by equation (11).

$r(c)=\sum_{s, m \in W} G_{s, m} \Omega_{s, m}(c)$ (11)

Conditions metV²(F) is given by equation (12):

$\int_{-\infty}^{+\infty}|r(c)|^2 q c<+\infty, c \in F$ (12)

The scaling function expression used in the roll-based fast character recognition algorithm is given by Eq. (13):

$\theta(c)=\left\{\begin{array}{l}1,0 \leq c \leq 1 \\ 0, \text { other }\end{array}\right.$ (13)

It uses the wavelet functionHalvalić as shown in the equation. (14):

$\Omega(c)=\theta(2 c)+\theta(2 c-1)$ (14)

Suppose the number of unit labels is expressed asJ, the value of the scale line at the right end is expressed asX, the value of the left scale line is expressed asI. The formula for calculating the scale value of the ith line of the minor scale in the appropriate scale unit is (15).

$S(i)=\frac{i-1}{j} \times(X-Y)+Y$ (15)

The reading recognition algorithm based on the deflection of the pointer mainly relies on the accurate calculation of the angle of the pointer to obtain the correct reading. While this method allows for relatively accurate identification in many cases, it still has some drawbacks. For example, perspective distortion can occur when the angle between the camera and the dial changes. This distortion will affect the calculation of the pointer angle and therefore the accuracy of the reading recognition. In the case of a multi-pointer energy meter, the method must simultaneously handle multiple viewing angles. This may increase the complexity of the recognition algorithm and may lead to recognition errors. The method of using the scale line value as a reference value to identify the reading has the advantages of reducing angle calculation errors, improving stability and greater adaptability to different gauges. In the application scenario in this article, it can support multi-point meters.

Based on the existing scale mark calibration, the reading corresponding to each scale line can be further determined. Suppose the number of detected ticks is expressed asI, distance to point (C_N,G_Rice) to the indicator lineoak+I+G=0 meansQ_N, the indicator reading is expressed asF_I, the readings corresponding to the upper and lower scale lines closest to the pointer are expressed asF_OtherIF_V, the distance between the top and bottom ticks closest to the pointer is usedQ₁IQ₀The distance between the pointer and the scale line can be calculated according to the equation. (16).

$q_n=\frac{\left|O c_n+U g_n+G\right|}{\sqrt{X^2+Y^2}} ; n=1,2 \cdots i$ (16)

The formula for calculating the meter reading is given by the equation. (17).

$F_i=F_V+\frac{q_1}{q_0+q_1} \times\lijevo(F_E-F_V\desno)$ (17)

Observing the binary horizontal projection of the image from the energy monitor shown in Figure 4, it can be seen that there may be obvious fluctuations in the brightness of some areas. These fluctuations can be caused by uneven lighting, reflections from instrument materials, and other factors. In this case, relying solely on binary processing may not be able to fully extract effective information from the instrument, which affects the subsequent identification and analysis process. Therefore, in order to obtain more accurate readings from energy monitoring instruments, it is necessary to adjust the brightness of the image, eliminate the influence of uneven lighting, and increase the readability and accuracy of image recognition.

Table 2 compares the results of the experiments before and after image preprocessing. A more detailed analysis of differences in experimental results before and after pretreatment can be done in four aspects: error analysis, time analysis, stability analysis, and practical analysis. The table shows that the recognition error after preprocessing is generally low. This shows that image pre-processing, such as brightness adjustment and Hough transform, can reduce uneven lighting and ambient noise in an image and make the instrument's display more recognizable. This helps to improve the accuracy of pointer detection and angle calculation, reducing recognition errors. The recognition time after pre-processing is generally shorter, which may be because the pre-processing process reduces noise and noise in the image, making it easier for the algorithm to recognize accurate features. In addition, the processed image is clearer, which helps the algorithm to quickly locate the hands and dial, thus reducing the recognition time. In the experimental results after pretreatment, the error and time fluctuation is small, which shows that the pretreatment method has better stability for instruments with different displayed values. This is critical for intelligent energy monitoring analysis, where stability is a key factor in practical applications where a large number of meters with different displayed values ​​need to be processed. The pre-treatment method not only has small error and short recognition time, but also has better adaptability to instruments with different displayed values. This makes the method more practical for practical applications and helps improve the efficiency and accuracy of intelligent energy monitoring and analysis. In conclusion, image preprocessing such as brightness adjustment and Hough transform have a significant impact on improving the recognition accuracy of energy monitoring instrument readings, reducing recognition time and increasing stability. In a real-world smart energy monitoring and analysis scenario, this pre-processing method is of great importance and helps to achieve efficient and accurate energy monitoring.

Figure 4.Horizontal projection of a binary image of an energy monitor

Table 2.Comparison of experimental results before and after image pre-processing

Figure 5.Run-time curves for different methods of pointer detection and pointer angle calculation

Figure 6.The result of fitting the straight line of the indicator

Figure 5 shows the time curves of different pointer detection and pointer angle calculation methods. By observing the data in the table, you can see that the operating time of each indicator detection method varies with different values ​​on the display of the instrument. The operation time of the method proposed in this article is relatively short at various values ​​displayed by the instrument, and the fluctuation is small, which proves the stability of the method and the high speed of calculations. PeriodostroumanThe method is subject to large fluctuations depending on the different values ​​displayed on the instrument, and the total operating time is much longer than other methods. It showsostroumanthe method may be less efficient in the task of detecting pointers. The uptime of the template matching method is relatively stable, but for some of the values ​​displayed by the gauge, the uptime is high. This suggests that template matching methods may be less effective in some cases. PeriodAuxiliary vector machinesThe method varies slightly depending on the different values ​​displayed on the instrument, but its total running time is longer than the proposed method. It showsAuxiliary vector machinesthe method is less efficient in the task of detecting the pointer. In short, according to the analysis of the operating time, the method proposed in the article is characterized by greater efficiency and stability in the task of detecting the indicator at various values ​​of the device's indications, and the efficiency is better thanostrouman, template matching andAuxiliary vector machinesmethod.

Figure 6 shows the results of fitting the indicator to a straight line. Observing this result, it can be seen that the straight line of the indicator skeleton is very close to the actual direction of the indicator, and the adjustment effect is better. In this work, the least squares method was used to fit the skeleton of the hand, the goal is to obtain a straight line, the sum of the distances from the straight line to the points of the hand skeleton is as small as possible. This method can effectively eliminate the effect of image noise on the editing result.

In addition, in this study, the experimental results of various methods of recognizing indications of energy meters were compared, and the results of the comparison are presented in Table 3. Comparing the methods proposed in this study,Maska R-CNN, Iwatershed, four conclusions can be drawn: error analysis, time analysis, stability analysis, and practical analysis. The table shows that the recognition error of the method proposed in this study is generally low, indicating that the recognition accuracy is better. mistakeMaska R-CNNIwatershedThese methods are relatively large, which may be due to their sensitivity to lighting and environmental disturbances when processing energy meter images. The recognition time of the method proposed in this study is generally shorter, which implies higher efficiency of processing images from energy meters. But the time of recognitionMaska R-CNNIwatershedThe method is long and may affect the efficiency of the intelligent energy monitoring analysis. In the experimental results of the method proposed in this paper, errors and time fluctuations are smaller, which indicates better stability. fluctuationsMaska R-CNNIwatershedThe method is relatively large and may cause unstable recognition results for instruments with different displayed values. The method proposed in this study has better adaptability in the case of meters with different displayed values, while maintaining a low error and short recognition time.Maska R-CNNIwatershedThe method performs poorly in terms of adaptability and may require fine-tuning of parameters for different instruments. To sum up, the method proposed in this paper is characterized by greater accuracy, shorter recognition time and better stability of recognizing electricity meter indications and a greater value of practical applicability than traditional methods.Maska R-CNNIwatershedmethod. In a real scenario of smart energy regulation analysis, efficient and accurate energy monitoring can be achieved by adopting the method proposed in this paper.

Finally, this paper compares the relative errors of different methods of recognizing electricity meter indications, and the results of the comparison are presented in Table 4.Maska R-CNN,watershed, before and after pre-treatment, the following conclusions can be drawn. Comparing the errors before and after pre-processing, it is clear that pre-processing significantly reduces the recognition error, which indicates the importance of the pre-processing step for improving the accuracy of recognizing electricity meter readings. The preprocessed method has less error compared toMaska R-CNNIwatershedmethods, proving that preprocessing has better performance in terms of recognition accuracy. In contrast,Maska R-CNNIwatershedThe error of the method is relatively large, which may be affected by factors such as light and environmental interference. Regarding error performance, the preprocessed method has less jitter and shows better stability.Maska R-CNNIwatershedThe method varies greatly, which can lead to unstable recognition results for instruments with different displayed values. The preprocessing method has better adaptability for instruments with different display values, and generally has a small error.Maska R-CNNIwatershedThe method performs poorly in terms of adaptability and may require fine-tuning of parameters for different instruments. To sum up, the method proposed in this paper, after initial processing, is characterized by greater accuracy and better stability in recognizing electricity meter indications and has a greater practical value than traditional methods.Maska R-CNNIwatershedmethod. In a real scenario of smart energy regulation analysis, efficient and accurate energy monitoring can be achieved by adopting the method proposed in this paper.

Table 3.Comparison of experimental results of different methods of recognizing indications of energy monitoring devices

Table 4.Comparison of relative errors of various methods of recognizing electricity meter readings