Substation infrared image mouse identification method based on deep learning

Abstract

The invention discloses a transformer substation infrared image mouse identification method based on deep learning, and relates to the technical field of deep learning, and the method comprises the steps: collecting mouse activity data in different time periods through a thermal wave radar infrared imaging technology; based on the mouse activity data, performing feature extraction on infrared images and radar data in the mouse activity data, and establishing feature vectors of the mouse; on the basis of the feature vectors of the rats, feature category labeling is conducted on the rats, grouping is conducted according to a preset proportion, learning training is conducted after training parameters are configured for a YOLOv5 model, and an optimal model is selected; the optimal model is put into use, a detection result output by the optimal model contains a confidence score, and when it is detected that the confidence score of the mouse is higher than a preset threshold value, an alarm is automatically triggered, and the ultrasonic mouse repeller is automatically started. The method has the advantages that the rat repelling speed and accuracy are improved, the time blank and missing detection possibility of manual inspection are avoided, and the safety of an indoor transformer substation is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of deep learning, and specifically relates to a method for identifying mice in substation infrared images based on deep learning. Background Art

[0002] The substation is a key hub of the power system and plays a crucial role. It can achieve voltage conversion, ensure the efficient transmission and reasonable distribution of electric energy, and meet the electricity demand of each region. At the same time, its functions of power dispatching control and fault protection maintain the stable operation of the power grid and are an indispensable infrastructure for the normal operation of modern society.

[0003] The harm of mice to substations is not only chewing cables, causing line short circuits and tripping, affecting power transmission, and even causing power outages, seriously interfering with production and living order; traditional substation inspections rely on manual inspections. Manually searching for mice is not only time-consuming and laborious, but also prone to omissions. In large indoor substations, there are many devices and complex environments. Manual inspections are usually carried out regularly and cannot achieve real-time monitoring of mouse activities. Summary of the Invention

[0004] To solve the above technical problems, a method for identifying mice in substation infrared images based on deep learning is provided, and this technical solution solves the above problems.

[0005] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0006] A method for identifying mice in substation infrared images based on deep learning, characterized by including:

[0007] Collect mouse activity data at different time periods through thermal wave radar infrared imaging technology;

[0008] Based on the mouse activity data, extract features from the infrared images and radar data therein to establish a feature vector of the mouse;

[0009] Based on the feature vector of the mouse, perform feature category annotation on it, group it according to a preset ratio, configure the training parameters of the YOLOv5 model and then perform learning and training to select the optimal model;

[0010] Put the optimal model into use. The detection result output by the optimal model has a confidence score. When the confidence score of detecting a mouse is higher than a preset threshold, an alarm is automatically triggered and an ultrasonic mouse repellent is automatically started.

[0011] Preferably, according to the method for identifying mice in substation infrared images based on deep learning described in claim 1, the step of collecting mouse activity data at different time periods through thermal wave radar infrared imaging technology specifically includes:

[0012] For the compact and narrow space layout of indoor substations, a thermal wave radar that can accurately detect small targets and has high range resolution and velocity resolution, and a high-resolution and high-sensitivity infrared thermal imager are selected. The installation positions of the two need to be opposite to each other to keep the viewing angles of the two overlapping.

[0013] Based on the relative positions of the thermal wave radar and the infrared thermal imager, the whole day is divided into working periods, lunch break periods, upper midnight periods, and lower midnight periods. Different scanning modes are adopted by the thermal wave radar and the infrared imager in different periods to collect mouse activity data.

[0014] Preferably, for a method for identifying mice in infrared images of substations based on deep learning according to claim 2, it is characterized in that, based on the mouse activity data, feature extraction is performed on the infrared images and radar data therein, and the establishment of the feature vector of the mouse specifically includes:

[0015] Extracting the infrared image features in the mouse activity data, including body temperature features and shape features, to obtain the average body temperature feature, the standard deviation of body temperature distribution, perimeter, aspect ratio, and circularity of the mouse;

[0016] Extracting the radar data features in the mouse activity data, including motion features, and obtaining the motion speed of the mouse through the radar measurement principle;

[0017] Based on the infrared image features and radar data features in the mouse activity data, a feature vector is established.

[0018] Preferably, for a method for identifying mice in infrared images of substations based on deep learning according to claim 3, it is characterized in that the extraction of the infrared image features in the mouse activity data, including body temperature features and shape features, specifically includes:

[0019] Calculation of the average body temperature feature: In the infrared image, determine the area where the mouse is located. For all pixel points in this area, obtain their temperature values and calculate the average value of these temperature values;

[0020] Assume that the area in the infrared image contains n pixel points, and their temperature values are T1, T2,..., T n , then the average body temperature calculation formula:

[0021]

[0022] is expressed as the average body temperature, T i is expressed as the temperature value of the i-th pixel point in the area where the mouse is located;

[0023] Calculation of the standard deviation of body temperature distribution: In addition to the average body temperature, calculate the standard deviation of the mouse's body temperature distribution within its body area to reflect the uniformity of body temperature;

[0024] Calculation formula:

[0025]

[0026] σ represents the standard deviation of body temperature distribution, m represents the number of pixel points within the mouse's body area, and T j represents the temperature value of the jth pixel point, represents the average body temperature.

[0027] Preferably, in a method for identifying mice in substation infrared images based on deep learning according to claim 3, the method is characterized in that extracting the radar data features in the mouse activity data, including motion features, and obtaining the motion speed of the mouse through the radar measurement principle specifically includes:

[0028] According to the Doppler effect, when the target moves relative to the radar, the frequency of the reflected wave will change, and this frequency change is proportional to the radial motion speed of the target. The radial motion speed of the mouse is calculated by the following formula:

[0029]

[0030] v r represents the radial motion speed of the mouse, λ represents the wavelength of the radar transmitted wave, and f d is the Doppler frequency shift, which is the difference between the received frequency and the transmitted frequency.

[0031] Preferably, in a method for identifying mice in substation infrared images based on deep learning according to claim 3, the method is characterized in that the method for calculating the feature vector:

[0032] Combine all the extracted features together to form the feature vector of the mouse, according to the extracted average body temperature, standard deviation of body temperature distribution, perimeter, aspect ratio, circularity, and motion speed;

[0033] The formula for the feature vector is expressed as:

[0034]

[0035] represents the feature vector, T avg represents the average body temperature, σT represents the standard deviation of body temperature distribution, P represents the perimeter, represents the aspect ratio, R represents the circularity, and v represents the motion speed.

[0036] Preferably, for a method for identifying mice in infrared images of a substation based on deep learning according to claim 3, it is characterized in that, based on the feature vectors of the mice, performing feature category annotation on them, grouping them according to a preset ratio, configuring training parameters for the YOLOv5 model and then performing learning and training, and selecting the optimal model specifically includes:

[0037] Performing feature grouping annotation on the extracted mouse feature vectors, and dividing the labeled data set into a training set, a validation set, and a test set according to a preset ratio;

[0038] Using the training set to train the YOLOv5 model, evaluating the model during the training process through the validation set, selecting the model with the highest average precision as the preferred model, and performing a final test on the preferred model on the test set to obtain the optimal model.

[0039] Preferably, for a method for identifying mice in infrared images of a substation based on deep learning according to claim 7, it is characterized in that, when putting the optimal model into use, the detection result output by the optimal model has a confidence score, and when the confidence score of the detected data is higher than a preset threshold, automatically triggering an alarm and automatically starting an ultrasonic mouse repeller specifically includes:

[0040] After triggering the alarm, the system automatically starts the ultrasonic mouse repeller, and the device emits ultrasonic waves at a specific frequency, which causes discomfort to the auditory system of the mice and forces the mice to flee the area.

[0041] Preferably, for a method for identifying mice in infrared images of a substation based on deep learning according to claim 8, it is characterized in that the detection result output by the optimal model has a confidence score specifically includes:

[0042] Performing normalization processing on the output result of the optimal model to convert it into a probability distribution representing the target belonging to each category;

[0043] The calculation formula for the classification probability is:

[0044]

[0045] P ab is the probability that the a-th prediction box belongs to the b-th category, z ab is the original score of the a-th prediction box output by the model for the b-th category, and C is the total number of target categories;

[0046] Calculation of intersection over union: Measuring the accuracy of the prediction box position by calculating the intersection over union of the prediction box and the ground truth box; the calculation formula for the intersection over union is:

[0047]

[0048] The IOU is expressed as the intersection over union, and the IOU is used as the location confidence; A1 is expressed as the area of the predicted bounding box, B1 is expressed as the area of the ground truth bounding box, and C1 is expressed as the intersection area of the predicted bounding box and the ground truth bounding box.

[0049] Multiply the classification probability and the location confidence to obtain the confidence score. The calculation formula is:

[0050] confidence = P mouse × P loc

[0051] P mouse represents the classification probability, that is, P mouse = P ab ; P loc represents the location confidence, that is, P loc = IOU.

[0052] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0053] The present invention proposes a method for identifying mice in infrared images of substations based on deep learning. Through the thermal wave radar infrared imaging technology, the activities of mice can be monitored throughout the entire time period. By using the deep learning method, the position, size, and species of mice can be accurately analyzed. Using the optimal model selected through training, when a mouse appears, an alarm can be automatically triggered to turn on the ultrasonic mouse repellent, and the mouse can be driven away by the specific frequency of the ultrasonic wave. In this way: the speed and accuracy of mouse repelling are improved, the time gap and the possibility of missed inspection in manual inspection are avoided, and the safety of indoor substations is enhanced. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 is a flowchart of a method for identifying mice in infrared images of substations based on deep learning;

[0055] Figure 2 is a flowchart of a method for collecting mouse activity data at different time periods;

[0056] Figure 3 is a flowchart of a method for extracting features of infrared images and radar data and establishing a feature vector of mice;

[0057] Figure 4 is a flowchart of a method for selecting the optimal model;

[0058] Figure 5 is a flowchart of a method for automatically triggering an alarm when a mouse is detected. DETAILED DESCRIPTION OF THE INVENTION

[0059] The following description is used to disclose the present invention so that those skilled in the art can implement the present invention. The preferred embodiments described below are only examples, and those skilled in the art can think of other obvious variations.

[0060] Referring to Figure 1 as shown, a method for identifying mice in substation infrared images based on deep learning includes:

[0061] Collect mouse activity data at different time periods through thermal wave radar infrared imaging technology;

[0062] Based on the mouse activity data, extract features from the infrared images and radar data therein, and establish a feature vector of the mouse;

[0063] Based on the feature vector of the mouse, perform feature category annotation on it, group it according to a preset ratio, configure the training parameters of the YOLOv5 model and then perform learning and training to select the optimal model;

[0064] Put the optimal model into use. The detection result output by the optimal model has a confidence score. When the confidence score of detecting a mouse is higher than a preset threshold, an alarm is automatically triggered and an ultrasonic mouse repellent is automatically started.

[0065] This solution can monitor the activity of mice throughout the whole time period through thermal wave radar infrared imaging technology, accurately analyze the position, size and species of mice through deep learning methods, use the optimal model selected by training, and can automatically trigger an alarm to turn on the ultrasonic mouse repellent when a mouse appears, and drive the mouse away through the specific frequency of the ultrasonic wave. In this way: the mouse repellent efficiency is greatly improved, the speed and accuracy of mouse repellent are improved, the time blind area of manual inspection and the possibility of missed inspection are avoided, and the safety of indoor substations is improved.

[0066] Referring to Figure 2 as shown, collect mouse activity data at different time periods through thermal wave radar infrared imaging technology, including:

[0067] For the compact and narrow space layout of indoor substations, select a thermal wave radar that can accurately detect small targets and has high range resolution and velocity resolution, and an infrared thermal imager with high resolution and high sensitivity. The installation positions of the two need to be opposite to each other to keep the viewing angles of the two overlapping;

[0068] Based on the relative positions of the thermal wave radar and the infrared thermal imager, divide the whole day into working hours, lunch break hours, first half of the night hours and second half of the night hours. Different scanning modes are used by the thermal wave radar and the infrared imager at different time periods to collect mouse activity data.

[0069] This solution collects data on the activities of mice around the clock through a thermal wave radar and an infrared thermal imager. The thermal wave radar and the infrared thermal imager are installed in relative positions, and their viewing angles overlap to ensure that there are no blind spots in the collection range. The analysis of the mouse activity data requires the data collected by both, ensuring the validity of the data analysis. In this way: collect mouse activity data around the clock, avoid monitoring blind spots and time gaps, and provide a rich basis for data analysis.

[0070] Refer to Figure 3 As shown, based on the mouse activity data, feature extraction is performed on the infrared images and radar data therein, and the establishment of the feature vector of the mouse includes:

[0071] Extract the infrared image features in the mouse activity data, including body temperature features and shape features, and obtain the average body temperature feature, the standard deviation of body temperature distribution, perimeter, aspect ratio, and circularity of the mouse;

[0072] Extract the radar data features in the mouse activity data, including motion features, and obtain the motion speed of the mouse through the radar measurement principle;

[0073] Based on the infrared image features and radar data features in the mouse activity data, establish a feature vector.

[0074] Assume that the area in the infrared image contains n pixel points, and their temperature values are T1, T2, …, T n , then the formula for calculating the average body temperature:

[0075]

[0076] Denoted as the average body temperature, T i Denoted as the temperature value of the i-th pixel point in the area where the mouse is located;

[0077] Calculation of the standard deviation of body temperature distribution: In addition to the average body temperature, calculate the standard deviation of the mouse body temperature distribution within its body area to reflect the uniformity of body temperature;

[0078] The calculation formula:

[0079]

[0080] σ is denoted as the standard deviation of body temperature distribution, m is denoted as the number of pixel points in the mouse body area, T j Denoted as the temperature value of the j-th pixel point, Denoted as the average body temperature.

[0081] According to the Doppler effect, when the target moves relative to the radar, the frequency of the reflected wave will change, and this frequency change is proportional to the radial motion speed of the target. The radial motion speed of the mouse is calculated by the following formula:

[0082]

[0083] v r represents the radial movement speed of the mouse, λ represents the wavelength of the radar transmitted wave, f d is the Doppler frequency shift, which is the difference between the received frequency and the transmitted frequency.

[0084] Combine all the extracted features together to form the feature vector of the mouse. According to the extracted average body temperature, standard deviation of body temperature distribution, perimeter, aspect ratio, circularity, and movement speed;

[0085] The calculation formula of the feature vector is expressed as:

[0086]

[0087] is expressed as the feature vector, T avg is expressed as the average body temperature, σT is expressed as the standard deviation of body temperature distribution, P is expressed as the perimeter, is expressed as the aspect ratio, R is expressed as the circularity, and v is expressed as the movement speed.

[0088] It should be noted that the calculation of the average body temperature feature: In the infrared image, determine the area where the mouse is located. For all pixel points in this area, obtain their temperature values and calculate the average of these temperature values.

[0089] This solution calculates and analyzes the mouse activity data to establish a feature vector. Extract the infrared image features and radar data features in the mouse activity data, obtain the average body temperature feature, standard deviation of body temperature distribution, perimeter, aspect ratio, and circularity of the mouse, and obtain the movement speed of the mouse. In this way: According to the characteristics of the mouse itself, as the judgment criterion, refine and quantify its characteristics.

[0090] Refer to Figure 4 As shown, based on the feature vector of the mouse, perform feature category annotation on it, group it according to a preset ratio, configure the training parameters of the YOLOv5 model and then perform learning and training, and select the optimal model including:

[0091] Perform feature grouping annotation on the extracted mouse feature vector, and divide the labeled data group into a training set, a validation set, and a test set according to a preset ratio;

[0092] Use the training set to train the YOLOv5 model, evaluate the model during the training process through the validation set, select the model with the highest average precision as the preferred model, and perform the final test of the preferred model on the test set to obtain the optimal model.

[0093] In this solution, the optimal model is selected by debugging and training the YOLOv5 model. Configure the training parameters, group and label the feature vectors, and then divide them into a training set, a validation set, and a test set according to a ratio to train and screen the YOLOv5 model to obtain the optimal model. In this way: complete the training and learning of the model and obtain the optimal model.

[0094] Refer to Figure 5 As shown, put the optimal model into use. The detection results output by the optimal model carry confidence scores. When the confidence score of the detected data is higher than the preset threshold, an alarm is automatically triggered and an ultrasonic rat repeller is automatically started, including:

[0095] After the alarm is triggered, the system automatically starts the ultrasonic rat repeller, which will emit ultrasonic waves of a specific frequency, causing discomfort to the auditory system of the rats and forcing the rats to flee the area.

[0096] In this solution, data with a confidence score higher than the preset value is identified as a rat, and the alarm and the ultrasonic rat repeller are automatically triggered. The functions of accurately identifying rats, automatically triggering alarms, and quickly repelling rats are realized.

[0097] The usage process of the present invention is as follows:

[0098] Step 1: For the compact and narrow space layout of the indoor substation, select a thermal wave radar that can accurately detect small targets and has high range resolution and velocity resolution, and an infrared thermal imager with high resolution and high sensitivity. The installation positions of the two need to be opposite to keep the viewing angles of the two overlapping;

[0099] Step 2: Based on the relative positions of the thermal wave radar and the infrared thermal imager, divide the whole day into a working period, a lunch break period, a first half of the night period, and a second half of the night period. Different scanning modes are adopted by the thermal wave radar and the infrared imager in different periods to collect rat activity data;

[0100] Step 3: Extract the infrared image features in the rat activity data, including body temperature features and shape features, to obtain the average body temperature feature, the standard deviation of body temperature distribution, the perimeter, the aspect ratio, and the circularity of the rats;

[0101] Step 4: Extract the radar data features in the rat activity data, including motion features, and obtain the motion speed of the rats through the radar measurement principle;

[0102] Step 5: Based on the infrared image features and radar data features in the rat activity data, establish a feature vector;

[0103] Step 6: Group and label the extracted rat feature vectors, and divide the labeled data groups into a training set, a validation set, and a test set according to a preset ratio;

[0104] Step 7: Train the YOLOv5 model using the training set, evaluate the model during the training process through the validation set, select the model with the highest average precision as the preferred model, and conduct the final test of the preferred model on the test set to obtain the optimal model;

[0105] Step 1: Put the optimal model into use. The detection results output by the optimal model carry confidence scores. When the confidence score of the detected data is higher than the preset threshold, an alarm is automatically triggered.

[0106] Step 8: After the alarm is triggered, the system automatically activates the ultrasonic rat repeller. This device emits ultrasonic waves at a specific frequency, which causes discomfort to the auditory system of the rats and forces the rats to flee the area.

[0107] In summary, the advantages of the present invention are as follows: The activities of rats can be monitored throughout the entire time period. The position, size, and species of rats can be accurately analyzed through deep learning methods. When rats appear, an alarm can be automatically triggered to turn on the ultrasonic rat repeller, and the rats can be expelled by the specific frequency of the ultrasonic waves, improving the speed and accuracy of rat expulsion, avoiding the time gap of manual inspection and the possibility of missed detection, and enhancing the safety of indoor substations.

[0108] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art of this industry should understand that the present invention is not limited by the above embodiments. What is described in the above embodiments and the specification is only the principle of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. A substation infrared image rat recognition method based on deep learning, characterized in that: include: Through thermal wave radar infrared imaging technology, the activity data of mice in different time periods are collected; Based on the mouse activity data, feature extraction is performed on the infrared image and radar data therein to establish a feature vector of the mouse; Based on the feature vector of the mouse, the feature category is marked, the mouse is grouped according to a preset ratio, the YOLOv5 model is configured with training parameters, and learning training is performed to select the optimal model; The optimal model is put into use, and the detection result output by the optimal model has a confidence score. When the confidence score of the detected mouse is higher than the preset threshold, an alarm is automatically triggered and the ultrasonic mouse repellent is automatically started.

2. According to the deep learning-based substation infrared image mouse recognition method of claim 1, it is characterized in that: The above-mentioned method of collecting rat activity data in different time periods by using thermal wave radar infrared imaging technology specifically includes: In view of the compact and narrow space layout of indoor substations, thermal wave radars with high distance resolution and speed resolution that can accurately detect tiny targets and infrared thermal imagers with high resolution and high sensitivity are selected. The two must be installed in opposite positions to keep their viewing angles overlapping; Based on the relative positions of the thermal wave radar and infrared thermal imager, the whole day is divided into working hours, lunch break, first half of the night and second half of the night. The thermal wave radar and infrared imager use different scanning modes to collect mouse activity data in different time periods.

3. A substation infrared image mouse recognition method based on deep learning according to claim 2, characterized in that: The extracting features of the infrared image and radar data based on the mouse activity data to establish the mouse feature vector specifically includes: Extract infrared image features from mouse activity data, including body temperature features and shape features, and obtain the average body temperature features, standard deviation of body temperature distribution, circumference, aspect ratio, and circularity of mice; Extract radar data features from mouse activity data, including motion features, and obtain the mouse's motion speed through radar measurement principles; Based on the infrared image features and radar data features in the mouse activity data, a feature vector is established.

4. A method for identifying rats in infrared images of substations based on deep learning according to claim 3, characterized in that: The infrared image features extracted from the mouse activity data, including body temperature features and shape features, specifically include: Calculation of average body temperature characteristics: In the infrared image, determine the area where the mouse is located, obtain the temperature values ​​of all pixels in the area, and calculate the average value of these temperature values; Assume that the area in the infrared image contains n pixels, whose temperature values ​​are T1, T2, …, T n , then the average body temperature calculation formula is: Expressed as average body temperature, T i It is represented as the temperature value of the i-th pixel in the area where the mouse is located; Calculation of the standard deviation of body temperature distribution: In addition to the average body temperature, the standard deviation of the distribution of the mouse's body temperature in its body area is calculated to reflect the uniformity of the body temperature; Calculation formula: σ represents the standard deviation of body temperature distribution, m represents the number of pixels in the mouse body area, and T j Represented as the temperature value of the jth pixel, Expressed as mean body temperature.

5. The method for identifying rats in infrared images of substations based on deep learning according to claim 3 is characterized in that: The extraction of radar data features from mouse activity data, including motion features, and obtaining the mouse's motion speed through radar measurement principles specifically include: According to the Doppler effect, when the target moves relative to the radar, the frequency of the reflected wave will change. This frequency change is proportional to the radial speed of the target. The radial speed of the mouse is calculated by the following formula: v r is the radial velocity of the mouse, λ is the wavelength of the radar emission wave, and f d is the Doppler shift, which is the difference between the received frequency and the transmitted frequency.

6. The method for identifying rats in infrared images of substations based on deep learning according to claim 3 is characterized in that: The method for establishing the feature vector calculation is as follows: All extracted features were combined to form a feature vector of the mouse, based on the extracted mean body temperature, standard deviation of body temperature distribution, circumference, aspect ratio, circularity, and movement speed; The eigenvector calculation formula is expressed as: Expressed as a feature vector, T avg It is represented by the average body temperature, σT is represented by the standard deviation of body temperature distribution, and P is represented by the circumference. It is expressed as the aspect ratio, R is expressed as the circularity, and v is expressed as the movement speed.

7. The method for identifying rats in infrared images of substations based on deep learning according to claim 3 is characterized in that: The method of labeling the mouse's feature vector based on the mouse's feature category, grouping the mouse according to a preset ratio, configuring the YOLOv5 model with training parameters, and performing learning and training to select the optimal model specifically includes: The extracted mouse feature vectors are grouped and labeled, and the labeled data set is divided into a training set, a validation set, and a test set according to a preset ratio; The YOLOv5 model is trained using the training set, and the model in the training process is evaluated using the validation set. The model with the highest average accuracy is selected as the preferred model, and the preferred model is finally tested on the test set to obtain the optimal model.

8. The method for identifying rats in infrared images of substations based on deep learning according to claim 7 is characterized in that: The optimal model is put into use, and the detection result output by the optimal model has a confidence score. When the confidence score of the detected data is higher than a preset threshold, an alarm is automatically triggered and the ultrasonic mouse repellent is automatically started. Specifically, the steps include: When the alarm is triggered, the system automatically activates the ultrasonic mouse repeller, which emits ultrasonic waves of a specific frequency that cause discomfort to the mouse's auditory system, forcing the mouse to flee the area.

9. A method for identifying rats in infrared images of substations based on deep learning according to claim 8, characterized in that: The detection results output by the optimal model with confidence scores specifically include: Normalize the output of the optimal model and convert it into a probability distribution indicating that the target belongs to each category; The calculation formula for classification probability is: P ab is the probability that the a-th prediction box belongs to the b-th category, z ab is the original score of the a-th prediction box output by the model for the b-th category, and C is the total number of target categories; Intersection-over-union calculation: The accuracy of the predicted box position is measured by calculating the intersection-over-union ratio between the predicted box and the true box; The calculation formula of intersection-over-union ratio is: IOU is represented by intersection over union, and IOU is used as position confidence; A1 is represented by the area of ​​the predicted box, B1 is represented by the area of ​​the real box, and C1 is represented by the intersection area of ​​the predicted box and the real box; Multiply the classification probability and the position confidence to get the confidence score. The calculation formula is: confidence=P mouse ×P loc P mouse Expressed as classification probability, that is, P mouse =P ab ;P loc Expressed as position confidence, that is, P loc =IOU.

Images