A digital instrument character detection method based on SW-FSSD

Abstract

The present invention discloses a digital instrument character detection method based on SW‑FSSD, which comprises: constructing a SW‑FSSD model as a digital instrument character detection model and performing model training; using IPC to collect character images of digital instruments; inputting character images into the trained digital instrument character detection model; using the digital instrument character detection model to extract and classify features of the character images, fusing the features of the high and low layers of the model to obtain character categories, and outputting character detection results. The present invention can improve the accuracy and effectiveness of digital instrument character detection and meet the needs of intelligent detection of digital instrument characters in traction substations.

Description

Technical Field

[0001] The invention relates to a digital instrument character detection method based on SW-FSSD, belonging to the technical field of digital instrument character detection. Background Art

[0002] The digital tube of the digital instrument contains an anode and a cathode made of metal wire, which are filled with different rare gases. With the help of electrodes, the power is supplied to emit different colors of light, thereby displaying relevant character data. Due to its low cost and good stability, digital instruments are widely used in the power industry to cumulatively record power data such as voltage and current, and to monitor and display the voltage and temperature information of substations and power supply stations in real time to ensure the safe and stable operation of the power system. However, digital instruments can only display power data, and then the displayed character data needs to be read and re-recorded manually, which is slow and easy to introduce human errors. Social development is gradually leaning towards intelligent power systems and new energy equipment, and the degree of dependence on electricity is becoming increasingly prominent. The digital instrument characters recorded manually obviously cannot meet the growing data in the power industry. Through the acquisition of real-time images of digital instrument characters and intelligent image recognition, the automatic reading of values ​​will be the development trend of future power monitoring systems, and the accuracy and effectiveness of instrument character recognition will further affect the development of traction substations towards automation, integration, visualization, and intelligence. Therefore, a reliable and stable instrument character recognition method is urgently needed.

[0003] The detection of digital instrument character values ​​belongs to a direction of character detection. As a research hotspot in the field of computer vision, traditional machine learning methods for character detection mainly use image processing technology. Most of the characters studied have simple backgrounds and are not easily affected by the environment. However, digital instrument characters are easily affected by factors such as background character imprints, uneven luminescence, and illumination. Using traditional image processing methods, manually designing and extracting features is complex and time-consuming, with poor robustness, and it is difficult to guarantee detection accuracy and speed. In recent years, with the development of society, the emergence of artificial intelligence technology, and the improvement of computer hardware level, the digital instrument character detection method based on deep learning uses end-to-end convolutional neural networks to automatically learn features. Compared with the staged image processing method, the detection speed has been greatly improved, and because the convolutional neural network is used to automatically obtain target features, various defects of manually designed features are avoided, and better performance is obtained. However, the method based on deep learning still has problems in the detection of digital instrument characters. The main reason is that the semantic information extracted by the backbone network feature is not rich enough and the number of samples is easily imbalanced when collecting samples, which ultimately reduces the performance of the model. The SSD model is a representative and excellent network model in deep learning networks. However, it still has the aforementioned problems when detecting digital instrument characters, which limits the accuracy of digital instrument character detection. Summary of the invention

[0004] In order to solve the problems existing in the prior art, the present invention proposes a digital instrument character detection method based on SW-FSSD, which is improved by multiple deep learning models SSD, introduces high- and low-level semantic information, and performs model training according to partial weighted classification loss, which can effectively improve the accuracy and effectiveness of digital instrument character detection.

[0005] In order to solve the above technical problems, the present invention adopts the following technical means:

[0006] The present invention proposes a digital instrument character detection method based on SW-FSSD, comprising the following steps:

[0007] Use IPC to collect character images of digital instruments;

[0008] Inputting the character image into a pre-built digital instrument character detection model, wherein the digital instrument character detection model adopts a SW-FSSD model;

[0009] The digital instrument character detection model is used to extract and classify the features of the character image, obtain the character category, and output the character detection result.

[0010] Furthermore, the SW-FSSD model includes a backbone feature extraction network, a multi-scale feature extraction network, a feature fusion structure and a classifier connected in sequence, wherein the backbone feature extraction network adopts the first thirteen convolutional layers of the VGG16 network, the multi-scale feature extraction network includes 3 groups of convolutional layers with outputs of different scale resolutions, and the feature fusion structure adopts the FPN structure.

[0011] Furthermore, the classification loss function of the SW-FSSD model adopts a partial weighted loss function SWLoss.

[0012] Furthermore, the training method of the digital instrument character detection model includes:

[0013] Acquire multiple character images of digital instruments;

[0014] Divide the character images into training sets and test sets, and annotate each character image;

[0015] Initialize the model parameters of the digital instrument character detection model;

[0016] The character image in the training set is processed by feature extraction, feature fusion and feature classification through the digital instrument character detection model to obtain the character image detection result;

[0017] According to the labeling and detection results of the character image, the total loss of the digital instrument character detection model is calculated using the loss function;

[0018] The model parameters of the digital instrument character detection model are updated according to the total loss, and the updated digital instrument character detection model is used to process the character images in the training set;

[0019] The model parameters are updated repeatedly until the total loss of the digital instrument character detection model converges, thereby obtaining a trained digital instrument character detection model.

[0020] Furthermore, the method for feature extraction, feature fusion and feature classification of the character images in the training set is:

[0021] Input the character images in the training set into the backbone feature extraction network, and obtain the first semantic feature through feature extraction;

[0022] Inputting the first semantic feature into a multi-scale feature extraction network, and obtaining the second, third, and fourth semantic features through feature extraction;

[0023] The feature fusion structure is used to fuse the four semantic features output by the backbone feature extraction network and the multi-scale feature extraction network to obtain a feature map with high- and low-level semantic information.

[0024] The classifier is used to classify the feature map to obtain the character category corresponding to the character image.

[0025] Furthermore, the total loss of the digital instrument character detection model includes classification loss and regression loss. The calculation formula of the total loss is as follows:

[0026]

[0027] Where L(x,c,l,g) represents the total loss of the digital instrument character detection model, x represents the indicator parameter of whether the predicted anchor box and the labeled anchor box match any character category, c represents the probability value of the character category predicted by the digital instrument character detection model, l represents the anchor box position predicted by the digital instrument character detection model, g represents the labeled anchor box position, κ is the multi-task imbalance factor, SWL conf (x,c) represents the partial weighted classification loss of the digital instrument character detection model, N is the total number of anchor boxes predicted by the digital instrument character detection model, and L loc (x,l,g) represents the regression loss of the digital instrument character detection model.

[0028] Furthermore, the partial weighted classification loss SWL conf The expression for (x,c) is as follows:

[0029]

[0030] Where p is the total number of character categories, k = 0, 1, ..., p, L conf(x,c) represents the classification loss for each character category, β k Represents the class imbalance factor for the k-th character class.

[0031] Furthermore, L conf The expression of (x,c) is as follows:

[0032]

[0033] Among them, Pos represents a positive sample, which refers to a predicted anchor box whose overlap with the labeled anchor box reaches a preset threshold; Neg represents a negative sample, which refers to a predicted anchor box whose overlap with the labeled anchor box does not reach a preset threshold. Indicates the indicator parameter of whether the i-th predicted anchor box matches the j-th labeled anchor box regarding the k-th character category, express The regularization value of represents the probability value of the i-th predicted anchor box being predicted as the k-th character category, express The regularization value of It represents the probability value that the i-th predicted anchor box is predicted as the background class, i = 1, 2, ..., N, j = 1, 2, ..., M, M is the total number of annotated anchor boxes in the character image.

[0034] Furthermore, the regression loss L loc The expression of (x,l,g) is as follows:

[0035]

[0036] Among them, Pos represents the positive sample, cx, cy, w, and h represent the horizontal coordinate, vertical coordinate, width, and height of the center point of the anchor box, respectively. They represent the encoded values ​​of the center point abscissa, center point ordinate, width and height of the i-th predicted anchor box converted to the anchor box coordinate system, They respectively represent the encoded values ​​of the center point abscissa, center point ordinate, width and height of the j-th annotation anchor box converted into the anchor box coordinate system.

[0037] Furthermore, in, They represent the horizontal coordinate, vertical coordinate, width and height of the center point of the jth annotation anchor box respectively. They respectively represent the horizontal coordinate, vertical coordinate, width and height of the center point of the pre-divided grid corresponding to the i-th predicted anchor box.

[0038] The following advantages can be obtained by using the above technical means:

[0039] The present invention proposes a digital instrument character detection method based on SW-FSSD. For digital instrument characters of the same scale, the digital instrument character detection model in the method of the present invention can retain four layers of prediction output layers, saving unnecessary calculations and prediction outputs; at the same time, the mechanism of the model fusing high-level semantic feature information to the bottom layer enriches the semantic feature information obtained by each output layer, eliminates the interference of the background imprint of the digital instrument character, and can improve the prediction accuracy of each output layer, thereby improving the accuracy of digital instrument character detection. In the model training process, the method of the present invention performs model training based on the partial weighted loss function SWLoss, which can effectively solve the problem of low overall detection accuracy of the model caused by the imbalance of the number of category samples during training, and improve the model detection performance.

[0040] The method of the present invention utilizes an improved deep learning model to process digital instrument character images, which can accurately detect character categories and improve the accuracy and effectiveness of digital instrument character detection, which is of great significance for the development of traction substations towards automation, integration, visualization, and intelligence. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 A flowchart of a digital instrument character detection method based on SW-FSSD according to the present invention;

[0042] Figure 2 The following is a flow chart of training and detecting a digital instrument character detection model in an embodiment of the present invention;

[0043] Figure 3 A schematic diagram of a character image of a digital instrument in an embodiment of the present invention;

[0044] Figure 4 A network structure diagram of a digital instrument character detection model in an embodiment of the present invention;

[0045] Figure 5 Schematic diagram of a feature fusion structure in an embodiment of the present invention. DETAILED DESCRIPTION

[0046] The technical solution of the present invention is further described below in conjunction with the accompanying drawings:

[0047] The present invention proposes a digital instrument character detection method based on SW-FSSD, which mainly includes the early model construction and model training process and the later process of using the model to detect character images in real time. Figure 1 , 2 As shown, the specific steps include:

[0048] Step A: Improve the existing deep learning SSD model and construct a SW-FSSD model that integrates high- and low-level semantic features for use as a digital instrument character detection model.

[0049] like Figure 4 As shown in FIG. 1 , the SW-FSSD model in the present invention mainly includes a backbone feature extraction network, a multi-scale feature extraction network, a feature fusion structure and a classifier connected in sequence, wherein the backbone feature extraction network adopts the first thirteen convolutional layers of the VGG16 network, the multi-scale feature extraction network includes three groups of convolutional layers with different scale resolution outputs, and the feature fusion structure adopts the FPN structure. The specific structure is as follows Figure 5 As shown in the figure, the classifier uses a softmax classifier. The four convolution output layers of the backbone feature extraction network and the multi-scale feature extraction network can output semantic information of four different scales. The feature fusion structure uses the FPN idea to fuse the four semantic information, and then uses the classifier to detect the fused feature map.

[0050] Step B: In order to improve the accuracy of model prediction, model training is required. During the training process, the partial weighted loss function SWLoss is used as the classification loss function of the SW-FSSD model, and the classification loss used for model parameter training and updating is adaptively weighted, that is, the category loss function calculated for training set samples of different categories is assigned weights representing the number of category samples, and then the multi-task imbalance factor is introduced and the regression loss is combined as the total loss for updating the model weight parameters.

[0051] In an embodiment of the present invention, the training method of the SW-FSSD model (or digital instrument character detection model) includes:

[0052] Step B01: Use IPC (network camera) to collect images of digital instrument characters to obtain multiple character images of digital instruments, such as Figure 3 During the acquisition process, images should be collected from different angles, different lighting, different brightness of digital instrument characters and other natural factors to ensure the diversity of the collected images.

[0053] Step B02: Divide the character images into a training set and a test set in a ratio of 6:4, and annotate each character image.

[0054] In the embodiment of the present invention, 250 character images are collected, and the numbers of various samples in the training set and the test set after being divided proportionally are shown in Table 1 and Table 2.

[0055] Table 1 Statistics of the number of training set samples

[0056]

[0057] Table 2 Statistics of the number of test set samples

[0058]

[0059] According to Table 1 and Table 2, due to the influence of equipment operation, there is an obvious imbalance in the number of samples among the collected characters.

[0060] The position and category (0-9) of each digital instrument character are marked in each character image to ensure the robustness of the model that can be trained later.

[0061] Step B03: After the outline of the model is constructed, in order to speed up the training, the model parameters of the digital instrument character detection model need to be initialized to obtain an initial digital instrument character detection model.

[0062] In the embodiment of the present invention, Kaiming Gaussian is used for initialization so that the variance of the output of each convolutional layer is 1. The weight initialization method is shown in the following formula: Among them, a is the negative semi-axis slope of the activation function Relu, n l is the dimension of the input, that is, n l = convolution kernel side length 2 × Number of channels.

[0063] Step B04: Perform feature extraction, feature fusion and feature classification processing on the character images in the training set annotated in step B02 through a digital instrument character detection model to obtain a character image detection result.

[0064] (1) The character images in the training set are input into the backbone feature extraction network. The convolutional layer of the backbone feature extraction network extracts the feature information in the character image to obtain the first semantic feature.

[0065] (2) The first semantic feature is input into a multi-scale feature extraction network. The three convolutional layers in the multi-scale feature extraction network further extract feature information to obtain the second, third, and fourth semantic features.

[0066] (4) The feature fusion structure is used to fuse the four semantic features output by the backbone feature extraction network and the multi-scale feature extraction network to obtain a feature map with high- and low-level semantic information.

[0067] Different from the SSD model, the model of the present invention has four output predictions of different semantic feature sizes. The model upsamples the semantic feature map of the latter layer and merges it with the semantic feature map extracted from the previous layer, making full use of the advantages of the underlying feature map for easy target finding and the high-level feature map for easy positioning. The highest convolution output layer is doubled upsampled by linear interpolation, and the underlying convolution output layer is reduced in dimension by 1×1 convolution without changing the size of the feature map. Then, the semantic information of the two is fused to provide more global information for the underlying semantic features and improve the positioning accuracy. The fusion here is the splicing between channels, and then the fused result is convolved with a convolution kernel of size 3×3 and step size 1 to eliminate the aliasing effect, and the output of each layer of feature map is unified into 256 semantic feature map outputs, which contain rich semantic feature information, to the same number of fused semantic feature maps of different sizes.

[0068] (5) Use the classifier to classify the feature map and obtain the character category corresponding to the character image. The high-level semantic feature output layer of the network model is integrated with the low-level semantic feature output layer, and a four-layer output layer is used to detect instrument characters, enriching the network model's extraction of digital instrument character features and eliminating the interference of digital instrument character background imprints.

[0069] Step B05: Calculate the total loss of the digital instrument character detection model using a loss function based on the labeling and detection results of the character image.

[0070] In the method of the present invention, the total loss of the digital instrument character detection model includes classification loss and regression loss. The classification loss uses a partial weighted loss function SWLoss to adaptively weight the classification loss, that is, the category loss function calculated for the training set samples of different categories is assigned a weight representing the number of category samples. The SWLoss function can effectively balance the imbalance of category samples between the collected data.

[0071] The mathematical expression of the partial weighted classification loss SWLoss is as follows:

[0072]

[0073] Among them, SWL conf (x,c) represents the weighted classification loss of the digital instrument character detection model for all character categories, x represents the indicator parameter of whether the predicted anchor box and the labeled anchor box match any character category, c represents the probability value of the character category predicted by the digital instrument character detection model, p is the total number of character categories, k=0,1,…,p, when k=0, it represents the background class, when k≠0, it represents the specific character category, L conf (x,c) represents the classification loss for each character category, β kRepresents the class imbalance factor for the k-th character class.

[0074] Class imbalance factor β of SWLoss function k It is used to adjust the size of the loss weight of each sample, and then adjust the parameters of the corresponding category in the model. Here, its value is selected as the total number of each category in each batch of training samples.

[0075] When the overlap between a predicted anchor box and a labeled anchor box reaches a preset threshold, the predicted anchor box is a positive sample, otherwise the predicted anchor box is a negative sample. All anchor boxes predicted by the model can be divided into positive samples and negative samples according to the preset threshold. conf The expression for (x,c) is as follows:

[0076]

[0077] Among them, Pos represents positive samples, Neg represents negative samples, Indicates the indicator parameter of whether the i-th predicted anchor box matches the j-th labeled anchor box regarding the k-th character category, When the i-th predicted anchor box matches the j-th labeled anchor box with respect to the k-th character category, When not matching express The regularization value of represents the probability value of the i-th predicted anchor box being predicted as the k-th character category, express The regularization value of It represents the probability value that the i-th predicted anchor box is predicted as the background class i=1,2,…,N,j=1,2,…,M,N is the total number of anchor boxes predicted by the digital instrument character detection model, and M is the total number of annotated anchor boxes in the character image.

[0078] Regression loss L loc The expression of (x,l,g) is as follows:

[0079]

[0080] Among them, l represents the anchor box position predicted by the digital instrument character detection model, g represents the marked anchor box position, cx, cy, w, and h represent the horizontal coordinate, vertical coordinate, width, and height of the center point of the anchor box, respectively. They represent the encoded values ​​of the center point abscissa, center point ordinate, width and height of the i-th predicted anchor box converted to the anchor box coordinate system, respectively. They respectively represent the encoded values ​​of the center point abscissa, center point ordinate, width and height of the j-th annotation anchor box converted into the anchor box coordinate system.

[0081] in, They represent the horizontal coordinate, vertical coordinate, width and height of the center point of the jth annotation anchor box respectively. They respectively represent the horizontal coordinate, vertical coordinate, width and height of the center point of the pre-divided grid corresponding to the i-th predicted anchor box.

[0082] smooth L1 The mathematical expression of the function is:

[0083]

[0084] According to formulas (5) and (7), the total loss of the digital instrument character detection model is calculated as follows:

[0085]

[0086] Among them, L(x,c,l,g) represents the total loss of the digital instrument character detection model, and κ is the multi-task imbalance factor.

[0087] The multi-task imbalance factor κ of the total loss is selected as the total number of categories contained in each batch training sample, and then adjusted by multiples according to the size of the gap between the classification loss and the regression loss during the training sample. By introducing the multi-task imbalance factor and balancing the loss weights of the model classification task and regression task, the model performance can be improved.

[0088] Step B06: updating the model parameters of the digital instrument character detection model according to the total loss, and using the updated digital instrument character detection model to process the character images in the training set.

[0089] Step B07, repeat steps B04 to B06, continuously update the model parameters until the total loss of the digital instrument character detection model converges, and obtain the optimal model parameters, bring the optimal model parameters into the digital instrument character detection model, and obtain a trained digital instrument character detection model for model testing and subsequent character detection use.

[0090] Step B08: Use the test set to perform performance evaluation on the trained digital instrument character detection model to obtain test results for each character category.

[0091] Step C: In the real-time detection process, the character image of the digital instrument is collected in real time using the IPC.

[0092] Step D: input the character image collected in step C into the digital instrument character detection model trained in step B, use the digital instrument character detection model to extract features and classify the character image, obtain character categories, and output character detection results.

[0093] In order to verify the effect of the method of the present invention, the embodiment of the present invention provides the following comparative experiment, using the traditional SSD model, the improved FSSD model and the method of the present invention to detect the character images in the test set, wherein the FSSD model adopts a network structure consistent with the present invention, but does not use the SWLoss function for model training, and only uses the traditional loss function for model training.

[0094] The detection accuracy of the three methods is shown in the following table:

[0095] Table 3

[0096]

[0097] It can be seen from the above table that the accuracy of the method of the present invention is generally higher than that of the other two methods.

[0098] With regard to the existing technologies for digital instrument character detection, traditional machine learning has the problems of slow speed and low robustness, and the accuracy of emerging deep learning methods is easily limited by the imbalance of collected category samples and the lack of rich semantic feature extraction. The method of the present invention improves the existing deep learning methods according to the needs of digital instrument character detection, which can effectively solve the problems of the existing technologies, can more accurately identify character categories, improve the accuracy and effectiveness of digital instrument character detection, and further accelerate the development trend of intelligent and unmanned traction substations.

[0099] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A digital instrument character detection method based on SW-FSSD, characterized in that: The steps include: Use IPC to collect character images of digital instruments; Inputting the character image into a pre-built digital instrument character detection model, wherein the digital instrument character detection model adopts a SW-FSSD model; Use the digital instrument character detection model to extract and classify the character image features, obtain the character category, and output the character detection result; The SW-FSSD model includes a backbone feature extraction network, a multi-scale feature extraction network, a feature fusion structure and a classifier connected in sequence, wherein the backbone feature extraction network uses the first thirteen convolutional layers of the VGG16 network, the multi-scale feature extraction network includes three groups of convolutional layers with different scale resolution outputs, and the feature fusion structure uses the FPN structure; The classification loss function of the SW-FSSD model adopts the partially weighted classification loss function SWLoss. The expression of the partially weighted classification loss function SWLoss is as follows: ； in, represents the partial weighted classification loss of the digital instrument character detection model, N is the total number of anchor boxes predicted by the digital instrument character detection model, Indicates the indicator parameter of whether the predicted anchor box and the labeled anchor box match any character category. represents the probability value of the character category predicted by the digital instrument character detection model, p is the total number of character categories, , represents the classification loss for each character category, Represents the class imbalance factor for the k-th character class.

2. A digital instrument character detection method based on SW-FSSD according to claim 1, characterized in that: The training method of the digital instrument character detection model includes: Acquire multiple character images of digital instruments; Divide the character images into training sets and test sets, and annotate each character image; Initialize the model parameters of the digital instrument character detection model; The character image in the training set is processed by feature extraction, feature fusion and feature classification through the digital instrument character detection model to obtain the character image detection result; According to the labeling and detection results of the character image, the total loss of the digital instrument character detection model is calculated using the loss function; The model parameters of the digital instrument character detection model are updated according to the total loss, and the updated digital instrument character detection model is used to process the character images in the training set; The model parameters are updated repeatedly until the total loss of the digital instrument character detection model converges, thereby obtaining a trained digital instrument character detection model.

3. A digital instrument character detection method based on SW-FSSD according to claim 2, characterized in that: The method for feature extraction, feature fusion and feature classification of character images in the training set is: Input the character images in the training set into the backbone feature extraction network, and obtain the first semantic feature through feature extraction; Inputting the first semantic feature into a multi-scale feature extraction network, and obtaining the second, third, and fourth semantic features through feature extraction; The feature fusion structure is used to fuse the four semantic features output by the backbone feature extraction network and the multi-scale feature extraction network to obtain a feature map with high- and low-level semantic information. The classifier is used to classify the feature map to obtain the character category corresponding to the character image.

4. A digital instrument character detection method based on SW-FSSD according to claim 2, characterized in that: The total loss of the digital instrument character detection model includes classification loss and regression loss. The total loss is calculated as follows: ； in, represents the total loss of the digital instrument character detection model, l represents the anchor box position predicted by the digital instrument character detection model, g represents the annotated anchor box position, is the multi-task imbalance factor, represents the partial weighted classification loss of the digital instrument character detection model, N is the total number of anchor boxes predicted by the digital instrument character detection model, Represents the regression loss of the digital meter character detection model.

5. A digital instrument character detection method based on SW-FSSD according to claim 4, characterized in that: The expression is as follows: ； Among them, Pos represents a positive sample, which refers to a predicted anchor box whose overlap with the labeled anchor box reaches a preset threshold; Neg represents a negative sample, which refers to a predicted anchor box whose overlap with the labeled anchor box does not reach a preset threshold. Indicates the indicator parameter of whether the i-th predicted anchor box matches the j-th labeled anchor box regarding the k-th character category, , express The regularization value of represents the probability value of the i-th predicted anchor box being predicted as the k-th character category, express The regularization value of represents the probability value of the i-th predicted anchor box being predicted as the background class, , , M is the total number of annotated anchor boxes in the character image.

6. A digital instrument character detection method based on SW-FSSD according to claim 4, characterized in that: Regression Loss The expression is as follows: ； Among them, Pos represents the positive sample, cx, cy, w, and h represent the horizontal coordinate, vertical coordinate, width, and height of the center point of the anchor box, respectively. , They represent the encoded values ​​of the center point abscissa, center point ordinate, width and height of the i-th predicted anchor box converted to the anchor box coordinate system, respectively. , They respectively represent the encoded values ​​of the center point abscissa, center point ordinate, width and height of the j-th annotation anchor box converted into the anchor box coordinate system.

7. A digital instrument character detection method based on SW-FSSD according to claim 6, characterized in that: , , , ,in, , , , They represent the horizontal coordinate, vertical coordinate, width and height of the center point of the jth annotation anchor box respectively. , , , They respectively represent the horizontal coordinate, vertical coordinate, width and height of the center point of the pre-divided grid corresponding to the i-th predicted anchor box.

Images