Facial detection target and marking matching method, storage medium and processor

Abstract

The present invention provides a facial detection target and marking matching method, storage medium and processor, wherein the method includes: obtaining test set data; calling a detection algorithm model to detect the test set data, and obtaining the model detection facial feature pixel coordinate axis data; obtaining the marked facial feature pixel coordinate axis data obtained by marking the test set data; matching the model detection facial feature pixel coordinate axis data with the marked facial feature pixel coordinate axis data, calculating the intersection and union ratio, and obtaining the intersection and union ratio data set; judging the ability of the algorithm model detection result. Correctly judging and screening out the human head pixel coordinates of correct detection, repeated detection, wrong detection, and missed detection in the algorithm model inspection results, the ability of the algorithm model can be objectively evaluated; improving the training efficiency of the algorithm in the next step, and providing the training direction of the algorithm model in the next step; it can also be used for the completed algorithm model to provide help and evidence for its debugging of the optimal threshold.

Description

[Technical field]

[0001] The present invention relates to the field of facial recognition technology, and in particular to a facial detection target and marking matching method, a storage medium and a processor. [Background technology]

[0002] In recent years, artificial intelligence has developed rapidly. Applications based on artificial intelligence have been widely expanded in the fields of face recognition, intelligent tracking, image recognition, etc., and have made great progress and success. As we all know, the premise of artificial intelligence face recognition and face recognition tracking is to be able to extract facial features and then identify them. Whether the artificial intelligence algorithm can accurately detect the head in the image and identify the facial features plays a crucial role. In other words, an accurate head detection algorithm is the basic premise of face recognition and recognition tracking. At present, the method of effectively and intuitively testing the ability of the head detection algorithm is to use the same test set to compare the detection algorithm model results with the marking results. When the head detection algorithm model works with low confidence, there will be a large number of repeated detections of a head, and non-heads will be mistakenly detected as heads. [Summary of the invention]

[0003] The technical problem to be solved by the present invention is to provide a facial detection target and labeling matching method, storage medium and processor, which can correctly judge and filter out the pixel coordinates of human heads with correct detection, repeated detection, wrong detection and missed detection in the algorithm model inspection results. The ability of the algorithm model can be objectively evaluated, and the average IOU, detection rate, precision rate and recall rate indicators can be given. The training efficiency of the next step of the algorithm is improved, and the training direction of the algorithm model is given; at the same time, it can also act on the completed algorithm model to provide help and evidence for debugging the optimal threshold.

[0004] In order to solve the above technical problems, on the one hand, an embodiment of the present invention provides a facial detection target and marking matching method, comprising: obtaining test set data; calling a detection algorithm model to detect the test set data to obtain model detected facial feature pixel coordinate axis data; obtaining marked facial feature pixel coordinate axis data obtained by marking the test set data;

[0005] The pixel coordinate axis data of the facial features detected by the model is matched with the pixel coordinate axis data of the marked facial features, and the intersection and union ratio between the two is calculated to obtain an intersection and union ratio data set; based on the intersection and union ratio data set, the ability of the algorithm model detection results is judged.

[0006] Preferably, the ability of the algorithm model to detect results is judged based on the intersection-and-union ratio data set, including: processing the intersection-and-union ratio data, and calculating the number of false detections, missed detections, and repeated detections of the test set data by calling the detection algorithm model.

[0007] Preferably, obtaining the test set data includes: customizing rules for the test set data.

[0008] Preferably, obtaining the marked facial feature pixel coordinate axis data obtained by marking the test set data includes: obtaining information on the marking of the test set data, marking the head in the information, and recording the pixel coordinates of the head.

[0009] Preferably, calling the detection algorithm model to detect the test set data to obtain the model detected facial feature pixel coordinate axis data includes: calling the detection algorithm model to detect the test set data and marking the facial feature pixel coordinates of each head.

[0010] Preferably, calculating the number of false detections, missed detections, and repeated detections of the test set data by calling the detection algorithm model includes: n rectangular boxes with human heads are marked, and the algorithm model detects m rectangular boxes with human heads, then the n×m intersection-and-union ratio queues are:

[0011]

[0012] Where: IOU is the intersection-over-union ratio, the sequence number before IOU is the key value, the maximum intersection-over-union ratio value and its corresponding key value are taken from the intersection-over-union ratio queue, and the intersection-over-union ratio value is judged to be greater than the first intersection-over-union ratio threshold. If so, the model detection is correctly matched with the marking frame, and the correct matching sequence is put into the correct matching sequence to obtain the number of correct recognitions;

[0013] The intersection and union ratio data and their corresponding key values ​​that are put into the correct matching sequence are removed from the intersection and union ratio data set, and the key value A is i M j , i≤n, j≤m for cutting, and get the key value A of the marked sequence i , i≤n and the key value M detected by the algorithm model j , j≤m, the cyclic intersection and ratio sequence will contain A i , i≤n or M j , the key value of j≤m is taken out for judgment and removed from the intersection and union ratio queue at the same time. j , if the intersection and union ratio of the key value of j≤m is greater than the threshold value 2, then it is put into the repeated detection sequence to obtain the number of repeated detections;

[0014] If it contains A i , if the intersection-and-union ratio data of the key value of i≤n is greater than the second intersection-and-union ratio threshold, it is put into the false detection sequence to obtain the number of false detections.

[0015] Preferably, the rules for customizing the test set data include: customizing the length and width or resolution of the test data set.

[0016] Preferably, after recording the pixel coordinates of the head, the method further includes: recording the pixel coordinate information of the head in an XML file, and outputting an XML file for each picture of the test set data.

[0017] Preferably, after marking the facial feature pixel coordinates of each head, the method further includes: storing the facial feature pixel coordinate data of each head in a txt document, and outputting the detected pixel coordinates of each head.

[0018] Preferably, the first intersection-over-union ratio threshold is 0.3-0.4.

[0019] Preferably, the second intersection-over-union ratio threshold is 0.4-1.

[0020] On the other hand, an embodiment of the present invention provides a storage medium, wherein the storage medium includes a stored program, wherein the program executes the above-mentioned facial detection target and marking matching method when running.

[0021] On the other hand, an embodiment of the present invention provides a processor, which is used to run a program, wherein the program executes the above-mentioned facial detection target and marking matching method when running.

[0022] Compared with the existing technology, the above technical solution has the following advantages: for the marking and algorithm model detection matching method, it can correctly judge and screen out the pixel coordinates of human heads with correct detection, repeated detection, wrong detection, and missed detection in the algorithm model inspection results; it can objectively evaluate the capabilities of the algorithm model and provide the average IOU, detection rate, precision rate, and recall rate indicators; it can improve the training efficiency of the next step of the algorithm and provide the training direction of the algorithm model in the next step; it can also act on the completed algorithm model to provide help and evidence for its debugging of the optimal threshold.

Brief Description of the Drawings

[0023] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0024] Figure 1 It is a schematic diagram of the intersection and comparison principle in target detection in the prior art.

[0025] Figure 2 It is a flow chart of the facial detection target and marking matching method of the present invention. [Specific implementation method]

[0026] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0027] Figure 1 Schematic diagram of the intersection and comparison principle in target detection in the prior art. Figure 1 As shown in the figure, Intersection-over-Union (IOU), a concept used in target detection, is the overlap ratio between the generated candidate bounding box and the original marked bounding box, that is, the ratio of their intersection to their union. The ideal situation is complete overlap, that is, the ratio is 1.

[0028] Embodiment 1

[0029] Figure 2 FIG. 1 is a flow chart of the face detection target and marking matching method of the present invention. Figure 2 As shown, a facial detection target and marking matching method is characterized by comprising the steps of:

[0030] S11, obtain test set data;

[0031] S12, calling the detection algorithm model to detect the test set data, and obtaining the pixel coordinate axis data of the facial features detected by the model;

[0032] S13, obtaining the pixel coordinate axis data of the marked facial features obtained by marking the test set data;

[0033] S14, matching the facial feature pixel coordinate axis data detected by the model with the facial feature pixel coordinate axis data marked, calculating the intersection-and-union ratio between the two, and obtaining an intersection-and-union ratio data set;

[0034] S15. Determine the ability of the algorithm model to detect results based on the intersection-union ratio data set.

[0035] According to the intersection-and-union data set, the ability of the algorithm model to detect the results includes: processing the intersection-and-union data, and calculating the number of false positives, missed positives, and repeated positives of the test set data by calling the detection algorithm model. Obtaining the test set data includes: customizing the rules of the test set data. The rules of customizing the test set data include: customizing the length and width or resolution of the test data set. Obtaining the marked facial feature pixel coordinate axis data obtained by marking the test set data includes: obtaining the information of marking the test set data, marking the heads in the information, and recording the pixel coordinates of the heads. Calling the detection algorithm model to detect the test set data and obtaining the model-detected facial feature pixel coordinate axis data includes: calling the detection algorithm model to detect the test set data and marking the facial feature pixel coordinates of each head. Calculating the number of false positives, missed positives, and repeated positives of the test set data by calling the detection algorithm model includes: marking n rectangular boxes of heads, and the algorithm model detects m rectangular boxes of heads, then the n×m intersection-and-union queues are:

[0036]

[0037] Wherein: IOU is the intersection-over-union ratio, the sequence number before IOU is the key value, the maximum intersection-over-union ratio value and its corresponding key value are taken from the intersection-over-union ratio queue, and the intersection-over-union ratio value is judged whether it is greater than the first intersection-over-union ratio threshold. If so, the model detection is correctly matched with the marking frame, and the correct matching sequence is put into the correct matching sequence to obtain the number of correct identifications; the intersection-over-union ratio data put into the correct matching sequence and its corresponding key value are removed from the intersection-over-union ratio data set, and the key value A is calculated. i M j , i≤n, j≤m for cutting, and get the key value A of the marked sequence i , i≤n and the key value M detected by the algorithm model j , j≤m, the cyclic intersection and ratio sequence will contain A i , i≤n or M j , the key value of j≤m is taken out for judgment and removed from the intersection and union ratio queue at the same time. j , j≤m, the intersection and union ratio of the key value is greater than the threshold value 2, then put it into the repeated detection sequence to get the number of repeated detections; if it contains A i , if the intersection-and-union data of the key value of i≤n is greater than the second intersection-and-union threshold, it is put into the false detection series to obtain the number of false detections. After recording the pixel coordinates of the head, it also includes: using an xml file to record the pixel coordinate information of the head, and outputting an xml file for each picture of the test set data. After marking the facial feature pixel coordinates of each head, it also includes: using a txt document to store the facial feature pixel coordinate data of each head, and outputting the pixel coordinates of each head detected. The first intersection-and-union threshold is 0.3~0.4. The second intersection-and-union threshold is 0.4~1.

[0038] Embodiment 2

[0039] Facial detection may be head detection, facial feature detection, etc. This embodiment is described by taking head detection as an example.

[0040] Test data collection:

[0041] In the specific implementation, the image pixel is set to be greater than 640 pixels × 480 pixels, the main object in the image is a person, and the main people in the image are divided into intervals of 0 to 20 people, 20 to 50 people, 50 to 100 people, 100 to 200 people, etc., and the heads of people in the image are clearly visible. In the specific implementation, the image is collected by cutting the camera video.

[0042] Data acquisition:

[0043] Loop through the labeled information of each test set image, mark the head in the image, record its pixel coordinates and store the coordinate information in an XML file. Output an XML file for each image, and name the file after the image. Head labeling is to mark the head in the image with a rectangular box similar to a screenshot, and to determine the head coordinates, you only need to store the coordinates of the upper left corner of the rectangular box (xmin, ymin) and the lower right corner (xmax, ymax).

[0044] The detection algorithm model is called for head detection in each test set image, and the pixel coordinates of each head detected are output. Each image has a txt file to store the coordinate data, and the file name is named after the image. The algorithm model detects the head and stores the detected head with a rectangular box. To determine the head coordinates, only the upper left corner coordinate point (xmin, ymin) and the lower right corner coordinate point (xmax, ymax) of the rectangular box need to be stored.

[0045] Data processing:

[0046] Use the python script tool to read the same-named head coordinates xml file and the head coordinates txt file detected by the algorithm model. Customize the method to process the file data and obtain two lists, which store the (xmin, ymin, xmax, ymax) of the marking and the algorithm model detection respectively. And assign a key value to each rectangular coordinate in the list. Each rectangular coordinate of the head has an independent and unique key value that matches the coordinate value. For example, if there are n heads to be marked, the marking list is:

[0047] {A 1 :(xmin, ymin, xmax, ymax),

[0048] A 2 :(xmin, ymin, xmax, ymax), ......

[0049] A n :(xmin,ymin,xmax,ymax)},

[0050] For example, an element A in the list 1 :(xmin, ymin, xmax, ymax), where A 1 is the key value, separated from the coordinates by a colon. The algorithm model detection data has m heads, and the algorithm model detection data list is:

[0051] {M 1 :(xmin, ymin, xmax, ymax),

[0052] M 2 :(xmin, ymin, xmax, ymax), ......

[0053] M m :(xmin,ymin,xmax,ymax)}.

[0054] Traverse the two lists and call the IOU method to calculate the IOU. For example, if there are n head rectangles marked, and the algorithm model detects m head rectangles, traverse the two lists and calculate n×m IOU values. Take the key values ​​of the two lists and recombine them, assign the key value to the calculated IOU, and create a new list to store the IOU, such as:

[0055]

[0056] Custom method processing results:

[0057] In data processing, the marking results are matched with the detection results of the algorithm model in pairs to calculate the IOU. The IOU method is called to calculate the IOU by traversing the two lists. For example, if there are n head rectangles marked, and the algorithm model detects m head rectangles, n×m IOU values ​​are calculated after traversing the two lists. The key values ​​of the two lists are taken and recombined, and the calculated IOU is assigned a key value. At the same time, a new list is created to store the IOU, and the sequence of the full IOU is obtained. The key value and its IOU are processed according to the custom method to determine the number of heads correctly detected by the algorithm model, the number of false detections, the number of re-detections, and the number of missed detections.

[0058] Remove the combinations with IOU equal to 0 from the total IOU sequence, loop through the IOU list and take out the combination with the maximum IOU value. If the maximum value is the key value A 1 M 1 If the IOU is greater than 0.3, the number of correctly detected people is increased by 1, and then the keyword containing A1 and M 1 All combinations of are removed from the IOU list. If A 1 If the IOU with other combinations is greater than 0.3, the number of rechecks will be increased by 1, such as A 1 M 2 and A 1 M 3 If the IOU of A is greater than 0.3, the number of retests is increased by 2. i , i≤n, M j， If j≤m and the IOU with other combinations is greater than 0.3, the number of false positives increases by 1, such as A 2 M 1 and A 3 M 1 If the IOU of the model is greater than 0.3, the number of false positives is increased by 2. The total number of labeled people minus the number of people correctly detected by the model equals the number of missed positives. Then the maximum IOU combination is taken from the IOU list and the previous steps are repeated until the maximum IOU is less than 0.3.

[0059] The intersection and union ratio data and their corresponding key values ​​that are put into the correct matching sequence are removed from the intersection and union ratio data set, and the key value A is i M j , i≤n, j≤m for cutting, and get the key value A of the marked sequence i , i≤n and the key value M detected by the algorithm model j , j≤m, the cyclic intersection and ratio sequence will contain A i , i≤n or M j , the key value of j≤m is taken out for judgment and removed from the intersection and union ratio queue at the same time. j , j≤m, the intersection and union ratio of the key value is greater than the threshold value 2, then put it into the repeated detection sequence to get the number of repeated detections; if it contains A i , if the key value of i≤n has an I / O ratio data greater than the second I / O ratio threshold, it is put into the false detection sequence to obtain the number of false detections. In specific implementation, the first I / O ratio threshold is 0.3-0.4. The second I / O ratio threshold is 0.4-1. In specific implementation, the first I / O ratio threshold and the second I / O ratio threshold can be set according to different scenarios, and the first I / O ratio threshold and the second I / O ratio threshold are not limited here.

[0060] Get the key value corresponding to the maximum IOU and the key value containing its A i , i≤n, M j , j≤m is removed from the IOU sequence, and then the key value corresponding to the maximum IOU is taken out, and the previous step is repeated until the maximum IOU value is less than the first intersection-over-union ratio threshold, and the loop is terminated. The correct matching sequence, repeated detection sequence, and false detection sequence are obtained, and then the number of correctly matched sequences is subtracted from the total number of marked heads to get the number of missed detections.

[0061] Calculation indicators:

[0062] The purpose of this test is to evaluate the algorithm model detection results based on the labeling, and to determine whether the algorithm model detection results are correct based on the labeling results. The correct matching series, repeated detection series, false detection series, and missed detection series are obtained, and their average IOU, detection rate, precision rate, and recall rate are calculated according to the method. The average IOU is the average IOU of the correct model detection.

[0063] Detection rate = total number of model detections / total number of markings.

[0064] Precision rate = number of correctly detected people / total number of model detections.

[0065] Recall rate = number of correctly detected people / total number of labeled people.

[0066] Embodiment 3

[0067] An embodiment of the present invention further provides a storage medium, which includes a stored program, wherein the program executes the above-mentioned facial detection target and marking matching method process when running.

[0068] Optionally, in this embodiment, the storage medium may be configured to store program codes for executing the following facial detection target and marking matching method flow:

[0069] S11, obtain test set data;

[0070] S12, calling the detection algorithm model to detect the test set data, and obtaining the pixel coordinate axis data of the facial features detected by the model;

[0071] S13, obtaining the pixel coordinate axis data of the marked facial features obtained by marking the test set data;

[0072] S14, matching the facial feature pixel coordinate axis data detected by the model with the facial feature pixel coordinate axis data marked, calculating the intersection-and-union ratio between the two, and obtaining an intersection-and-union ratio data set;

[0073] S15. Determine the ability of the algorithm model to detect results based on the intersection-union ratio data set.

[0074] Optionally, in this embodiment, the above-mentioned storage medium may include but is not limited to: a USB flash drive, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk or an optical disk, and other media that can store program codes.

[0075] It can be seen that by adopting the storage medium of the present invention, the storage capacity is reduced, the program running speed of the built-in facial detection target and the marking matching method process is faster, the marking and the algorithm model detection are matched, and the pixel coordinates of the human head with correct detection, repeated detection, wrong detection, and missed detection in the algorithm model inspection results can be correctly judged and screened. The ability of the algorithm model can be objectively evaluated, and the average IOU, detection rate, precision rate, and recall rate indicators are given; the training efficiency of the next step of the algorithm is improved, and the training direction of the algorithm model is given. At the same time, it can also act on the completed algorithm model to provide help and evidence for its debugging of the optimal threshold.

[0076] Embodiment 4

[0077] An embodiment of the present invention further provides a processor, which is used to run a program, wherein the program executes the steps in the above-mentioned facial detection target and marking matching method when running.

[0078] Optionally, in this embodiment, the above program is used to perform the following steps:

[0079] S11, obtain test set data;

[0080] S12, calling the detection algorithm model to detect the test set data, and obtaining the pixel coordinate axis data of the facial features detected by the model;

[0081] S13, obtaining the pixel coordinate axis data of the marked facial features obtained by marking the test set data;

[0082] S14, matching the facial feature pixel coordinate axis data detected by the model with the facial feature pixel coordinate axis data marked, calculating the intersection-and-union ratio between the two, and obtaining an intersection-and-union ratio data set;

[0083] S15. Determine the ability of the algorithm model to detect results based on the intersection-union ratio data set.

[0084] Optionally, the specific examples in this embodiment may refer to the above embodiments and the examples described in the specific implementation, and this embodiment will not be described in detail here.

[0085] It can be seen that by adopting the processor of the present invention, the amount of data to be processed is reduced, the program running speed of the built-in facial detection target and the marking matching method process is faster, the marking and the algorithm model detection are matched, and the pixel coordinates of the human head with correct detection, repeated detection, wrong detection, and missed detection in the algorithm model inspection results can be correctly judged and screened. The ability of the algorithm model can be objectively evaluated, and the average IOU, detection rate, precision rate, and recall rate indicators are given; the training efficiency of the next step of the algorithm is improved, and the training direction of the algorithm model is given. At the same time, it can also act on the completed algorithm model to provide help and evidence for its debugging of the optimal threshold.

[0086] It can be seen from the above description that the facial detection target and labeling matching method, storage medium and processor according to the present invention can correctly judge and filter out the pixel coordinates of human heads with correct detection, repeated detection, wrong detection and missed detection in the algorithm model inspection results by using the labeling and algorithm model detection matching method. The ability of the algorithm model can be objectively evaluated, and the average IOU, detection rate, precision rate and recall rate indicators can be given; the training efficiency of the next step of the algorithm can be improved, and the training direction of the algorithm model can be given. At the same time, it can also act on the completed algorithm model to provide help and evidence for debugging the optimal threshold.

[0087] The embodiments of the present invention are described in detail above. Specific examples are used herein to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A facial detection target and labeling matching method, It is characterized in that include: Get the test set data; Call the detection algorithm model to detect the test set data and obtain the pixel coordinate axis data of the facial features detected by the model; Obtain the pixel coordinate axis data of the marked facial features obtained by marking the test set data; Match the pixel coordinate axis data of the facial features detected by the model with the pixel coordinate axis data of the marked facial features, calculate the intersection-and-union ratio between the two, and obtain the intersection-and-union ratio data set; According to the intersection-and-union ratio data set, judging the ability of the algorithm model to detect the results, the ability of judging the algorithm model to detect the results according to the intersection-and-union ratio data set includes: processing the intersection-and-union ratio data, calculating the number of false detections, the number of missed detections, and the number of repeated detections of the test set data by calling the detection algorithm model; The calculation of the number of false detections, missed detections, and repeated detections of the test set data by the detection algorithm model includes: n rectangular boxes marked with human heads, and the algorithm model detects m rectangular boxes with human heads, then the n×m intersection-and-union ratio queues are: Where: IOU is the intersection-over-union ratio, the sequence number before IOU is the key value, the maximum intersection-over-union ratio value and its corresponding key value are taken from the intersection-over-union ratio queue, and the intersection-over-union ratio value is judged to be greater than the first intersection-over-union ratio threshold. If so, the model detection is correctly matched with the marking frame, and the correct matching sequence is put into the correct matching sequence to obtain the number of correct recognitions; The intersection and union ratio data and their corresponding key values ​​that are put into the correct matching sequence are removed from the intersection and union ratio data set, and the key value A is i M j , i≤n, j≤m for cutting, and get the key value A of the marked sequence i , i≤n and the key value M detected by the algorithm model j , j≤m, the cyclic intersection and ratio sequence will contain A i , i≤n or M j , the key value of j≤m is taken out for judgment and removed from the intersection and union ratio queue at the same time. j , if the intersection and union ratio of the key value of j≤m is greater than the threshold value 2, then it is put into the repeated detection sequence to obtain the number of repeated detections; If it contains A i , if the intersection-and-union ratio data of the key value of i≤n is greater than the second intersection-and-union ratio threshold, it is put into the false detection sequence to obtain the number of false detections.

2. The facial detection target and marking matching method according to claim 1, It is characterized in that Obtaining test set data includes: customizing rules for test set data.

3. The facial detection target and marking matching method according to claim 1, It is characterized in that The pixel coordinate axis data of the marked facial features obtained by marking the test set data includes: Obtain information on labeling the test set data, mark the head in the information, and record the pixel coordinates of the head.

4. The facial detection target and marking matching method according to claim 1, It is characterized in that Call the detection algorithm model to detect the test set data, and obtain the facial feature pixel coordinate axis data detected by the model, including: The detection algorithm model is called to detect the test set data and mark the facial feature pixel coordinates of each head.

5. The facial detection target and marking matching method according to claim 2, It is characterized in that The rules for customizing test set data include: customizing the length and width or resolution of the test data set.

6. The facial detection target and marking matching method according to claim 3, It is characterized in that After recording the pixel coordinates of the head, it also includes: using an XML file to record the pixel coordinate information of the head, and outputting an XML file for each picture of the test set data.

7. The facial detection target and marking matching method according to claim 4, It is characterized in that After marking the pixel coordinates of the facial features of each head, it also includes: Use a txt file to store the facial feature pixel coordinate data of each head, and output the pixel coordinates of each head detected.

8. The facial detection target and marking matching method according to claim 1, It is characterized in that The first intersection-over-union ratio threshold is 0.3-0.

4.

9. The facial detection target and marking matching method according to claim 1, It is characterized in that The second intersection-over-union ratio threshold is 0.4-1.

10. A storage medium, It is characterized in that The storage medium includes a stored program, wherein the facial detection target and marking matching method according to any one of claims 1 to 9 is executed when the program is run.

11. A processor, It is characterized in that The processor is used to run a program, wherein the facial detection target and marking matching method according to any one of claims 1 to 9 is executed when the program is run.

Images