CN110781852A - A shoe-wearing sequence footprint recognition method based on multi-resolution feature fusion - Google Patents

Abstract

The invention provides a shoe wearing sequence footprint identification method based on multi-resolution feature fusion, which comprises the following steps: extracting a single shoe print relative pressure distribution image of an image to be identified; constructing a two-dimensional multi-resolution step energy diagram and a one-dimensional multi-resolution step energy diagram; extracting a footprint attitude offset angle from the single footprint two-dimensional multi-resolution step-method energy map binary group; grouping the footprint attitude offset angles by using a grouping model; reducing the identification range according to the similar images in the grouped sub-databases; and calculating the matching score of the multi-resolution step energy diagram to be recognized and the feature library so as to obtain a recognition result based on the footprint hierarchical recognition strategy. The method comprises the steps of segmenting a footprint sequence, then performing wavelet decomposition to construct a footprint attitude foot grouping model and a multi-resolution step energy map quadruplet, and performing identity recognition through integration of scores of three features based on the quadruplet and a footprint hierarchical recognition strategy.

Description

A shoe-wearing sequence footprint recognition method based on multi-resolution feature fusion

technical field

The invention relates to the technical field of footprint recognition, in particular, to a method for recognizing footprints of shoe-wearing sequences based on multi-resolution feature fusion.

Background technique

At present, biometric identification based on sequence footprint is mainly divided into sequence barefoot footprint recognition and sequence shoe footprint recognition. The commonly used method for sequence barefoot footprint recognition is to extract the centroid pressure trajectory features of the footprints during the walking process for identification ^[1] . But barefoot footprints don't fit the real-life scenario, so they're not very practical. For sequential shoe-wearing footprint recognition methods: (1) Obtain the footstep stride feature ^[2,3] , obtain the distance between the corresponding points of the front and rear adjacent footsteps (such as the edge point of the back heel) as the step length, and connect the centroid of the adjacent footsteps. The line is the walking line, the angle between the walking line and the center line of the footprint is the step angle, and the distance between the centroid of the footprint and the walking line on the opposite side is the step width. In this method based on stride information, the information obtained by quantitatively using the features of stride length, stride width and stride angle is unstable. The stride feature has little difference between different people, and can only be used as a narrowing range in the identification stage. use. (2) By constructing the average pressure distribution map ^[4] , calculating the normalized cross-correlation and the centroid offset angle to obtain the similarity score for identification. However, there are few studies on the sequence of wearing shoes, and the recognition accuracy of the existing recognition models based on wearing shoes is not high.

references:

[1] Zhou B, Singh M S, Doda Set al. The carpet knows: Identifying people in a smart environment from a single step. IEEE International Conference on Pervasive Computing and Communications Workshops, 2017.

[2] Lv Xinhua. A quantitative test method using the comprehensive index of stride characteristics [P]. Shandong: CN107527345A, 2017-12-29.

[3] Pan Nan, Wu Xing, Li Yan, Liu Yi. Design and implementation of intelligent stride feature analysis and inspection system [J]. Science and Technology and Engineering, 2014, 14(03): 64-69.

[4] Wang Xinnian, Chen Wenchao, Yu Dan, Wang Yaling. A method for identifying shoe footprint sequences based on pressure features [P]. Liaoning Province: CN110188694A, 2019-08-30.

SUMMARY OF THE INVENTION

According to the above-mentioned technical problem of low accuracy of shoe-wearing sequence footprint recognition, a method for shoe-wearing sequence footprint recognition based on multi-resolution feature fusion is provided. The invention performs wavelet decomposition after the footprint sequence is segmented to construct a footprint posture foot grouping model and a multi-resolution footwork energy map quadruple, and performs identity recognition through three feature score fusion based on quadruples and a footprint hierarchical identification strategy.

The technical means adopted in the present invention are as follows:

A shoe-wearing sequence footprint recognition method based on multi-resolution feature fusion, characterized in that it includes: an offline training process and an online identification process; the offline process at least includes the following steps:

Extracting the relative pressure distribution image of a single shoe print from the footprint images of the sequence to be trained, and splicing the single shoe print relative pressure distribution image group to construct a six-tuple of footprint expression;

Construct a 2D multi-resolution gait energy map;

Build a one-dimensional multi-resolution footwork energy map;

Extract the footprint attitude offset angle from the 2-tuple of the 2-D multi-resolution footwork energy map of a single footprint;

Use the above footprint attitude offset angle to build a footprint attitude grouping model;

Two-dimensional multi-resolution gait energy maps and one-dimensional multi-resolution gait energy maps feature libraries are constructed.

The online identification process includes at least the following steps:

Extracting a single shoe print relative pressure distribution image from the sequence footprint images to be identified, and splicing the single shoe print relative pressure distribution image group to construct a footprint expression six-tuple;

Construct a 2D multi-resolution gait energy map;

Build a one-dimensional multi-resolution footwork energy map;

Extract the footprint attitude offset angle from the 2-tuple of the 2-D multi-resolution footwork energy map of a single footprint;

Grouping the footprint posture offset angles by using a grouping model;

Narrow the recognition range according to the similar images in the sub-database after grouping;

The matching score between the energy map of the multi-resolution footwork to be recognized and the feature library is calculated to obtain the recognition result based on the hierarchical recognition strategy of footprints.

Compared with the prior art, the present invention has the following advantages:

(1) The present invention takes into account the different characteristics of footprint pressure images obtained at different resolutions, and performs wavelet decomposition on the images to obtain gait energy maps with different resolutions, thereby making the extracted information more complete.

(2) The present invention considers the footprint posture offset angles of the sole region and the heel region of the two-dimensional footwork energy map, and averages the footprint posture offset angles of the left and right feet under multi-resolution to obtain a more stable posture offset angle. , which can be used as a preliminary screening condition to quickly narrow down the range of identified persons.

(3) The present invention constructs a one-dimensional multi-resolution footwork energy map in order to reduce the interference of the shoe print pattern information during the identification of the shoe footprint image. The expression of the projection feature can remove the pattern influence and highlight the pressure distribution.

(4) The present invention extracts three feature matching scores under multi-resolution, can remove the influence of patterns, strengthen mutual constraints between features, and obtain more accurate and stable features.

(5) The present invention fuses the matching scores of the two-dimensional multi-resolution footwork energy map and the one-dimensional multi-resolution footwork energy map, and obtains a more accurate person identification result through a hierarchical recognition strategy of footprints.

To sum up, when applying the present invention, wavelet decomposition is used to obtain gait energy maps of different resolutions, combined with the offset angle of footprint posture under multi-resolution, and a more accurate personal identification result is obtained through the hierarchical recognition strategy of footprints, which solves the problem of existing problems. The characteristic information is unstable and the recognition efficiency is not high in the shoe-wearing sequence pressure footprint recognition technology.

Based on the above reasons, the present invention can be widely promoted in the fields of footprint sequence identification and the like.

Description of drawings

In order to illustrate the embodiments of the present invention or the technical solutions in the prior art more clearly, the following briefly introduces the accompanying drawings that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description These are some embodiments of the present invention, and for those of ordinary skill in the art, other drawings can also be obtained from these drawings without any creative effort.

FIG. 1 is a flowchart of the identification method of the present invention.

FIG. 2 is a schematic diagram of a footprint posture grouping model of the present invention.

FIG. 3 is a stitched image with a step size based on the left foot in an embodiment of the present invention.

FIG. 4 is a de-stepped stitched image based on the left foot in an embodiment of the present invention

FIG. 5 is a stitched image with step size based on the right foot in an embodiment of the present invention

FIG. 6 is a de-stepped stitched image based on the right foot in an embodiment of the present invention.

FIG. 7 is a line connecting the geometric centers of the sole area and the heel area in the embodiment of the present invention.

Detailed ways

In order to make those skilled in the art better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only Embodiments are part of the present invention, but not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

As shown in FIG. 1 , the present invention provides a method for identifying shoe-wearing sequence footprints based on multi-resolution feature fusion, which is characterized in that, comprising:

Step 1: Extract the relative pressure distribution image of a single shoe print in the image to be identified, and splicing the single shoe print relative pressure distribution image group to construct a footprint expression hexatuple.

Specifically, the original footprint sequence image is denoised, and the sequence image horizontal projection is segmented to obtain a left foot single shoe print relative pressure image group I ¹ and a right foot single shoe print relative pressure image group I ² . The relative pressure image group of a single shoe print is spliced. According to the principle of reducing error accumulation and the smallest unit that can reflect people's walking habits, the number of splicing feet is determined to be 2, and the benchmarks are splicing with step length and without step length. Splicing, the sequence relationship is "left and right" and "right left", so it can respectively form a stitched image group I ³ with a step size based on the left foot, a stitched image group I ⁴ with a step size based on the left foot, and the right foot is The reference image group I ⁵ with step size stitching and the right foot as the reference image group I ⁶ without step size stitching, as shown in Figure 3-6. The set I=(I ^k , k=1, 2, 3, 4, 5, 6) of the above six groups of images is called a footprint expression six-tuple.

Traverse the images under each tuple separately, find the maximum size of the image under each tuple as the standard size of the tuple S _Q1 , S _Q2 , S _Q3 , S _Q4 , S _Q5 , S _Q6 , and obtain each footprint image The circumscribed rectangle of , which is normalized to the standard size under the same tuple by zero-padding.

Step 2. Construct a two-dimensional multi-resolution footwork energy map.

表示每个人第k元足迹二维步法能量图，m表示每个人每个元组下的图片数量，构建每个人二维步法能量图六元组

足迹二维步法能量图六元组由单枚足迹二维步法能量图二元组和多枚足迹二维步法能量图四元组构成。Specifically, the normalized images are added and averaged by tuples to obtain a two-dimensional step energy map.

Indicates the 2D footstep energy map of the k-th element footprint of each person, m represents the number of pictures under each tuple of each person, and constructs a 6-tuple of each person's 2D footwork energy map

The 6-tuple of footprint 2D footwork energy map is composed of a single footprint 2D footwork energy map 2-tuple and multiple footprint 2D footwork energy map 4-tuple.

表示第k元第t个分辨率的足迹二维步法能量图。The two-dimensional step energy map of the footprint is subjected to B-layer wavelet decomposition, and B satisfies G represents the maximum width of the sole pattern, and the low-frequency coefficients of the last T layer of low resolution are extracted as the two-dimensional multi-resolution footwork energy map of the footprint, denoted as

Represents the 2D footstep energy map of the k-th element at the t-th resolution.

Step 3. Construct a one-dimensional multi-resolution footwork energy map.

提取投影向量

为第k元第t个二维步法能量图水平方向投影向量，

为垂直方向投影向量。Specifically, for the two-dimensional multi-resolution step method energy map

extract projection vector

is the horizontal projection vector of the k-th t-th two-dimensional step energy map,

is the vertical projection vector.

为二维多分辨率步法能量图的像素值，x,y分别表示能量图像素点的横坐标和纵坐标，H^(k,t)为第k元第t个能量图高度，W^(k,t)为第k元第t个能量图宽度。in,

is the pixel value of the two-dimensional multi-resolution step energy map, x, y represent the abscissa and ordinate of the energy map pixel respectively, H ^{(k, t)} is the k-th element t-th energy map height, W ^{(k ,t)} is the width of the t-th energy map of the k-th element.

为第k元第t个分辨率下的一维步法能量图。Construction of Footprint 1D Multi-Resolution Footwork Energy Map Quads from Projection Vectors

is the one-dimensional stepping energy map at the k-th element and the t-th resolution.

Step 4. Extract the footprint attitude offset angle from the two-tuple of the two-dimensional multi-resolution footwork energy map of a single footprint.

按3:2高度比例分为脚掌区域和脚跟区域，分别提取脚掌区域和脚跟区域的几何中心或质心，计算其连线与垂直方向的夹角θ^(k,t)，如图7所示。Specifically, the two-dimensional multi-resolution footwork energy map of a single footprint

According to the height ratio of 3:2, it is divided into the sole area and the heel area. The geometric center or centroid of the sole area and the heel area are extracted respectively, and the angle θ ^{(k, t)} between the connection line and the vertical direction is calculated, as shown in Figure 7.

其中和分别表示左、右脚脚印姿态偏移角。The average value of the angles calculated by the energy maps of each resolution under each tuple is taken as the footprint attitude angle of the tuple, namely

in and Represents the left and right footprint posture offset angles, respectively.

Step 5. Build a footprint attitude offset angle grouping model. Specifically include:

(a) Calculate the footprint attitude angle of each person in each tuple in the data set;

(b) Respectively count the maximum and minimum values of the left and right footprint posture angles in the data set, and determine the range of the values;

(c) Constructing a footprint pose grouping model: firstly divide the left footprint posture angle into L sub-intervals at equal intervals; then divide the obtained L sub-intervals into R sub-intervals at equal intervals according to the right footprint posture offset angle. L×R subintervals are obtained, as shown in Figure 2. In this embodiment, L is 2, and R is 4.

Step 6. Hierarchically group the footprint dataset. Specifically include:

(a) Construct a two-dimensional multi-resolution gait energy map of each person in the dataset according to the above method, denoted as

(b) Construct a one-dimensional multi-resolution gait energy map of each person in the dataset according to the above method, denoted as

(c) Calculate the footprint pose angle of each person in the dataset.

(d) According to the calculated footprint attitude angle and the above-mentioned footprint attitude grouping model, the entire footprint dataset is grouped according to the footprint attitude angle to form L×R sub-data sets.

Step 7: Calculate the matching score between the energy map of the multi-resolution footwork to be identified and the feature library, so as to obtain the identification result based on the hierarchical identification strategy of footprints.

Specifically, calculating the matching score between the multi-resolution footwork energy map to be identified and the feature library includes:

(a) Match score calculation based on footprint 2D multi-resolution footwork energy map

采用计算对应元组的归一化互相关得到相似度得分：2D multi-resolution footwork energy map for footprints

The similarity score is obtained by computing the normalized cross-correlation of the corresponding tuples:

代表待识别图像的二维多分辨率步法能量图，

代表了数据库中的二维多分辨率步法能量图，u,v分别表示能量图像素点横坐标和纵坐标的偏移量，

代表位于待识别能量图

覆盖下的区域的均值，r代表得到的互相关图。in,

a 2D multi-resolution gait energy map representing the image to be identified,

represents the two-dimensional multi-resolution step energy map in the database, u, v represent the offset of the abscissa and ordinate of the energy map pixel, respectively,

The representative is located in the energy map to be identified

The mean of the area under coverage, and r represents the resulting cross-correlation plot.

Then the matching score of the image to be recognized and the two-dimensional multi-resolution footwork energy map of the library map is:

W_2D＝[0.3 0.2 0.3 0.2]表示加权系数，也可以根据具体应用调整。in

W _2D =[0.3 0.2 0.3 0.2] represents the weighting coefficient, which can also be adjusted according to specific applications.

(b) Calculation of matching score based on footprint 1D multi-resolution footwork energy map

计算对应元组的投影距离：1D multi-resolution footwork energy map for footprints

Compute the projected distance of the corresponding tuple:

代表待识别图像的一维多分辨率步法能量图，

代表数据库中一维多分辨率步法能量图，dp^(i,k,t)为待识别图像和库图的投影距离。Among them, p is the p-norm, which can be 0.5 or 1, and can also be adjusted according to specific applications.

represents the one-dimensional multi-resolution gait energy map of the image to be recognized,

Represents the one-dimensional multi-resolution step energy map in the database, and dp ^(i,k,t) is the projected distance between the image to be recognized and the library map.

In order to fuse the similarity scores of each feature, the obtained distances are normalized to obtain the normalized similarity:

Then the matching score of the image to be recognized and the one-dimensional multi-resolution footwork energy map of the library map:

In this embodiment, W _1D =[0.3 0.15 0.5 0.05] represents a weighting coefficient, which can also be adjusted according to specific applications.

(c) Random forest matching score calculation based on footprint 2D multi-resolution footwork energy map

Scan the pixel values line by line for the two-dimensional multi-resolution step energy map to form a vector v ^(i,k,t) , i=1,2,...,N, where N is the total number of people in the database, and t is the t-th resolution , after the vector v ^(0,k,t) obtained from the step energy map to be identified enters the random forest model of the corresponding database, the probability sr ^(i,k,t) of each category in the database is obtained, i=1,2 ,…,N.

Then the two-dimensional multi-resolution footwork energy map random forest probability matching score of the image to be recognized and the library map is:

Wherein W _2DR =[0.3 0.2 0.4 0.1] represents the weighting coefficient, which can also be adjusted according to specific applications.

(d) Footprint multi-resolution footwork energy map matching score calculation

The three matching scores of the corresponding resolution are weighted and fused, and the weighting coefficient is determined through training, which is generally 0.4, 0.2, 0.4, and the matching score based on the multi-resolution footwork energy map is obtained:

Wherein W _SC =[0.4 0.2 0.4] represents a weighting coefficient, which can also be adjusted according to specific applications.

Further, the obtaining of the identification results based on the footprint hierarchical identification strategy includes:

(a) Determine the sub-data set where the footprints to be identified are located. After the footprints are to be identified to obtain the offset angle of the footprint, first determine the group l, l∈[1, L], and then determine the group r,r∈[1,R] where the offset angle of the right foot footprint posture is located, and finally determine the sub-data set where it is located, denoted as SD.

(b) Calculate the matching score of the footsteps to be identified with the footsteps of each person in the sub-dataset SD.

(c) Arrange the matching scores of the multi-resolution footwork energy maps in descending order, take the labels corresponding to the top two scores for each layer, and take the corresponding 2T labels for T resolutions. The identification result is determined by a majority voting mechanism among these 2T tags.

(d) If the majority vote fails, take the resolution with the top two scores in the T resolutions with a larger difference between the scores and a score difference greater than the threshold f ₀ , and the label corresponding to the highest score at this resolution is the recognition result, and the threshold _f0 is determined by training and is usually 0.2.

(e) If step (d) fails, compare the highest score among the T resolutions. If the highest score is greater than the threshold f ₁ , the label corresponding to the score is the recognition result, and the threshold f ₁ is determined by training, usually 0.5.

(f) If step (e) fails, then refuse to recognize the image.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, but not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: The technical solutions described in the foregoing embodiments can still be modified, or some or all of the technical features thereof can be equivalently replaced; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the technical solutions of the embodiments of the present invention. scope.

Claims (8)

1. A shoe wearing sequence footprint identification method based on multi-resolution feature fusion is characterized by comprising the following steps: an off-line training process and an on-line identification process;

the offline training process comprises at least the following steps:

extracting a single shoe print relative pressure distribution image from a sequence footprint image to be trained, and splicing the single shoe print relative pressure distribution image group to construct a footprint expression six-tuple;

constructing a two-dimensional multi-resolution step-method energy map;

constructing a one-dimensional multi-resolution step-method energy map;

extracting a footprint attitude offset angle from the single footprint two-dimensional multi-resolution step-method energy map binary group;

constructing a footprint attitude grouping model by using the footprint attitude offset angle;

constructing a two-dimensional multi-resolution step energy diagram and a one-dimensional multi-resolution step energy diagram feature library;

the online identification process comprises at least the following steps:

extracting a single shoe print relative pressure distribution image from the sequence footprint image to be recognized, and splicing the single shoe print relative pressure distribution image group to construct a footprint expression six-tuple;

constructing a two-dimensional multi-resolution step-method energy map;

constructing a one-dimensional multi-resolution step-method energy map;

extracting a footprint attitude offset angle from the single footprint two-dimensional multi-resolution step-method energy map binary group;

grouping the footprint attitude offset angles by using a grouping model;

reducing the identification range according to the similar images in the grouped sub-databases;

and calculating the matching score of the multi-resolution step energy diagram to be recognized and the feature library so as to obtain a recognition result based on the footprint hierarchical recognition strategy.

2. The method for identifying footmarks of sequences of wearing shoes based on multi-resolution feature fusion as claimed in claim 1, wherein the footmark expression six-tuple comprises a left foot single shoe print relative pressure image group I ¹Right single shoe print relative pressure image group I ²Splicing image group I with step length by taking left foot as reference ³Step-length-removing spliced image group I with left foot as reference ⁴Splicing image group I with step length by taking right foot as reference ⁵And step-length-removing splicing image group I taking right foot as reference ⁶。

3. The method for identifying footmarks of sequences of shoes worn based on multi-resolution feature fusion according to claim 1 or 2, wherein the constructing the two-dimensional multi-resolution step energy map comprises:

normalizing the image under each tuple to the standard size under the same tuple by a zero padding method, adding the normalized images according to the tuples to average to obtain a two-dimensional step method energy map, and constructing each person two-dimensional step method energy map six tuples;

after B-layer wavelet decomposition is carried out on the obtained footprint two-dimensional step method energy graph, the last T-layer low-frequency coefficient of low resolution is extracted to be used as the footprint two-dimensional multi-resolution step method energy graph and recorded as the footprint two-dimensional multi-resolution step method energy graph

A footprint two-dimensional step-by-step energy map representing the kth resolution of the kth element, wherein

G represents the maximum width of the tread.

4. The method for identifying footmarks of sequences of shoes worn based on multi-resolution feature fusion as claimed in claim 3, wherein the constructing the one-dimensional multi-resolution step energy map comprises:

for two-dimensional multi-resolution step-wise energy mapping

Extracting projection vectors

Wherein

A vector is projected in the horizontal direction of the t two-dimensional step energy diagram of the kth element,

projecting a vector for the kth two-dimensional step method energy diagram of the kth element in the vertical direction;

construction of footprint one-dimensional multi-resolution step-method energy map quadruplet according to projection vector

Wherein

Is a one-dimensional step method energy diagram under the t resolution of the kth element.

5. The method for recognizing footmarks of sequences of shoes worn based on multi-resolution feature fusion according to claim 4, wherein the step of calculating the footprint attitude angles comprises the steps of:

respectively extracting single footprint two-dimensional multi-resolution step-by-step energy map

The geometric centers or the mass centers of the sole area and the heel area are calculated, and the included angle theta between the connecting line of the geometric centers or the mass centers and the vertical direction is calculated ^(k,t)；

Taking the average value of the angles calculated by the resolution energy maps of each tuple as the footprint attitude angle of the tuple, namely

Wherein

And respectively representing the left and right footprint attitude angles.

6. The method for recognizing footmarks of sequences of wearing shoes based on multi-resolution feature fusion according to claim 5, wherein the step of constructing the grouping model comprises the following steps:

calculating the footprint attitude angle of each person under each tuple in the data set;

respectively counting the maximum value and the minimum value of the left and right footprint attitude angles in the data set, and determining the interval range of the foot print attitude angle value;

dividing the left footprint attitude angle into L sub-intervals at equal intervals; and dividing the obtained L sub-intervals into R sub-intervals at equal intervals according to the attitude angle of the right footprint, and finally obtaining the L multiplied by R sub-intervals.

7. The method for recognizing footmarks of sequences of wearing shoes based on multi-resolution feature fusion according to claim 1, wherein the calculating the matching scores of the multi-resolution gait energy map to be recognized and the feature library comprises:

calculating a matching score based on the footprint two-dimensional multi-resolution step-method energy map, a matching score based on the footprint one-dimensional multi-resolution step-method energy map and a random forest matching score based on the footprint two-dimensional multi-resolution step-method energy map;

and carrying out weighted fusion on the matching scores to obtain the matching scores based on the multi-resolution step method energy diagram and the feature library.

8. The method for recognizing footmarks of sequences of shoes worn based on multi-resolution feature fusion as claimed in claim 7, wherein the obtaining of recognition results based on the footmarks hierarchical recognition strategy comprises:

(a) determining a subdata set where the footprints to be identified are located and recording the subdata set as SD;

(b) calculating the matching score of the footprints to be recognized and the footprints of each person in the subdata set SD by the multi-resolution step method energy map;

(c) the matching scores of the multi-resolution step-method energy diagram are arranged in a descending order, the labels corresponding to the scores of the first two labels are taken from each layer, and the identification result is determined by adopting a majority voting mechanism for 2T labels on the T layer;

(d) if the majority table fails to pass, the score difference of the first two digits in the T resolution ratios is greater than the threshold value f ₀The label corresponding to the highest score under the resolution is the identification result;

(e) if step (d) does not pass, comparing the highest score of the T resolutions, and if the highest score is greater than the threshold f ₁If so, the label corresponding to the score is the identification result;

(f) if step (e) does not pass, rejecting the image for recognition.

Images