Dynamic adaptive weighted face recognition multi-algorithm hybrid fusion method

Abstract

The invention discloses a dynamic adaptive weighted face recognition multi-algorithm hybrid fusion method. The method comprises the following steps: S1, normalizing an evaluation system; s2, determining an adaptive weighting function; and S3, hybrid fusion: performing hybrid fusion on the similarity score output by each face recognition algorithm through two fusion modes of unary fusion and binary fusion. Through combination of evaluation system normalization, the adaptive weighting function and hybrid fusion, the accuracy of a single face recognition algorithm can be effectively improved, and the problems that a fusion strategy is too single and the recognition robustness is poor can be solved.

Description

Technical Field

[0001] The present invention relates to the technical field of face recognition, and in particular to a dynamic adaptive weighted face recognition multi-algorithm hybrid fusion method. Background Art

[0002] The current face recognition is mainly used to compare the target face with the specified face image (1:1), or compare the target face with the specified face database (1:N), and finally output the similarity score. The accuracy of face recognition is often improved by fusing the results of multiple face recognition algorithms. The existing face recognition multi-algorithm fusion method has the following disadvantages:

[0003] (1) The scores of different face recognition algorithms have different meanings. Some face recognition algorithms have a relatively concentrated score distribution, while others are relatively dispersed. The threshold selection is also different, and there is a lack of unified standards.

[0004] (2) The fusion method of the weighted fusion algorithm is too simple and does not take into account the weight of the score confidence. The higher the score confidence, the higher the weight value should be and the more reliable the result.

[0005] (3) The fusion strategy is too simplistic, resulting in poor recognition robustness. Summary of the invention

[0006] In view of the deficiencies in the prior art, the present invention aims to provide a dynamic adaptive weighted face recognition multi-algorithm hybrid fusion method.

[0007] In order to achieve the above object, the present invention adopts the following technical solution:

[0008] A dynamic adaptive weighted face recognition multi-algorithm hybrid fusion method comprises the following steps:

[0009] S1. Normalization of evaluation system: Normalize each face recognition algorithm according to the following formula:

[0010]

[0011] Among them, x i is the similarity score obtained by comparing the i-th set of face recognition algorithms, T i is the similarity threshold set by the i-th face recognition algorithm, s i is the similarity score obtained by the i-th face recognition algorithm after normalization through the evaluation system;

[0012] S2. Determine the adaptive weighting function:

[0013] The total interval of similarity score [0,100] is divided into three intervals, [0,t 1 ]、(t 1, t2 ),[t 2 ,100], where the threshold T∈(t 1, t 2 ), the adaptive weighting function is as follows:

[0014]

[0015] Among them, s i is the similarity score obtained by normalizing the evaluation system for the i-th face recognition algorithm, t 1 ,t 2 is the interval value corresponding to the i-th face recognition algorithm, r 0 is the weight value corresponding to the similarity threshold T of the i-th face recognition algorithm, that is, the minimum value of the weight function, r 2 t 2 The corresponding weight value, r 1 t 1 The corresponding weight value;

[0016] S3, Hybrid Fusion:

[0017] Through the two fusion methods of unary fusion and binary fusion, the similarity scores output by each face recognition algorithm are mixed and fused. The similarity score Output after mixed fusion is calculated as follows:

[0018]

[0019] In the above formula, A represents the result of the unary fusion strategy, and B represents the result of the binary fusion strategy. represents the weight of face recognition algorithm i in the unary fusion strategy, Indicates the weight of face recognition algorithm i in the binary fusion strategy. The same face recognition algorithm i is combined with different face recognition algorithms in the binary fusion strategy. The values ​​of are the same or different; and are the weights of the unary fusion strategy and the binary fusion strategy respectively; They represent the weights of the similarity scores output by face recognition algorithm i, calculated by formula (3);

[0020] f(s a ,s b ) is the new model formed by combining face recognition algorithm a and face recognition algorithm b in the binary fusion strategy, a,b=1,2,…,N and a≠b, is f(s a ,s b) is calculated by formula (3); a and b are non-repeating and disordered combinations of all face recognition algorithms participating in the hybrid fusion, and N represents the total number of face recognition algorithms participating in the hybrid fusion.

[0021] Furthermore, in step S3, there are:

[0022]

[0023] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and the computer program implements the above method when executed by a processor.

[0024] The present invention also provides a computer device, comprising a processor and a memory, wherein the memory is used to store a computer program; and when the processor is used to execute the computer program, the above method is implemented.

[0025] The beneficial effect of the present invention is that the present invention can effectively improve the accuracy of a single face recognition algorithm through the combination of evaluation system normalization, adaptive weighting function and hybrid fusion, and can solve the problem of overly simplistic fusion strategy and poor recognition robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 This is the overall framework diagram of the method of Example 1 of the present invention;

[0027] Figure 2 It is a normalized function graph of the evaluation system of the method of Example 1 of the present invention;

[0028] Figure 3 is a weight function graph of the method of embodiment 1 of the present invention;

[0029] Figure 4 Schematic diagram of the hybrid fusion process of the method in Example 1 of the present invention. DETAILED DESCRIPTION

[0030] The present invention will be further described below in conjunction with the accompanying drawings. It should be noted that this embodiment is based on the technical solution and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to this embodiment.

[0031] This embodiment provides a dynamic adaptive weighted face recognition multi-algorithm hybrid fusion method, which can improve face recognition accuracy. Figure 1 As shown, the method consists of three steps: normalizing the evaluation system, determining the adaptive weighting function, and hybrid fusion.

[0032] S1. Normalization of the evaluation system: To address the problem that the similarity scores output by different face recognition algorithms have different meanings, normalization of the evaluation system can reduce the differences in the score distribution of different face recognition algorithms. The specific calculation is as follows:

[0033]

[0034] Among them, x i is the similarity score obtained by comparing the i-th set of face recognition algorithms, T i is the similarity threshold set by the i-th face recognition algorithm, s i is the similarity score of the ith face recognition algorithm after normalization through the evaluation system. After normalization through the evaluation system, the threshold T of each face recognition algorithm will be set to 50. The function graph is as follows: Figure 2 shown.

[0035] S2. Determine the adaptive weighting function: In practical applications, the similarity score also affects the confirmation of the final result. When the similarity score is in the extremely low and extremely high ranges on both sides of 0 and 100, the credibility is higher, and when it is in the "ambiguous" range near the threshold of 50, the credibility is lower. Therefore, different weights should be used to describe this relationship according to the confidence of different similarity scores.

[0036] According to the actual application situation, this embodiment divides the total interval of similarity score [0,100] into three intervals: [0,t 1 ]、(t 1, t 2 ),[t 2 ,100], where the threshold T∈(t 1, t 2 ), these three intervals need to meet the following conditions:

[0037] (1) In the interval [0,t 1 ] and the interval [t 2 ,100] should be greater than the weight function value in the interval (t 1, t 2 ) is the weight function value within .

[0038] (2) In the interval [0,t 1 ] range, the weight function is monotonically decreasing.

[0039] (3) In the interval [t 2 ,100], the weight function is monotonically increasing.

[0040] (4) In the interval (t 1, t 2 ), the weight function is symmetric about the threshold T, and the threshold T corresponds to the minimum value of the weight function.

[0041] (5) The absolute value of the slope of the weight function near the threshold T is large, and the corresponding weight function value changes more dramatically, which can distinguish the "ambiguous" samples near the threshold T more clearly.

[0042] (6) For points far away from the threshold T, such as points close to 0 and close to 100, the weight function value should be very large, indicating that the confidence in the output of the face recognition algorithm is very high.

[0043] (7) At the point t where the intervals connect 1 ,t 2 Nearby, the weight function value changes relatively smoothly, ensuring that the weight function value does not change significantly near these points.

[0044] According to the above situation, this embodiment designs an adaptive weighting function If the above 7 conditions are met, the formula is as follows:

[0045]

[0046] Among them, s i is the similarity score obtained by normalizing the evaluation system for the i-th face recognition algorithm, t 1 ,t 2 is the interval value corresponding to the i-th face recognition algorithm, r 0 is the weight value corresponding to the similarity threshold T = 50 of the i-th face recognition algorithm, that is, the minimum value of the weight function, r 2 t 2 The corresponding weight value, r 1 t 1 The corresponding weight value, for different face recognition algorithms i, its t 1 ,t 2 、r 0 、r 2 The values ​​of may not be the same. The adaptive weighting function graph is as follows Figure 3 shown.

[0047] S3, Hybrid Fusion:

[0048] A single fusion strategy can easily lead to poor robustness of the multi-algorithm fusion model. This embodiment proposes a hybrid fusion strategy, which uses two fusion methods, unary fusion and binary fusion, to hybridly fuse the similarity scores output by each face recognition algorithm. Figure 4 As shown, the similarity score Output after hybrid fusion is calculated as follows:

[0049]

[0050] In the above formula, A represents the result of the unary fusion strategy, and B represents the result of the binary fusion strategy. represents the weight of face recognition algorithm i in the unary fusion strategy, Indicates the weight of face recognition algorithm i in the binary fusion strategy. The same face recognition algorithm i is combined with different face recognition algorithms in the binary fusion strategy. The value of can be different. and are the weights of the unary fusion strategy and the binary fusion strategy respectively. They respectively represent the weight values ​​of the similarity scores output by face recognition algorithm i, which are calculated by formula (3).

[0051] f(s a ,s b ) is a new model formed by combining face recognition algorithm a and face recognition algorithm b in the binary fusion strategy, a, b = 1, 2, ..., N and a ≠ b, and the mean is taken in this embodiment, is f(s a ,s b ) is calculated by formula (3); a and b are non-repeating and disordered combinations of all face recognition algorithms participating in the hybrid fusion, and N represents the total number of face recognition algorithms participating in the hybrid fusion.

[0052] Taking N=3 as an example, formula (4) can be expressed as:

[0053]

[0054] For face recognition algorithm 1, in the binary fusion strategy, in the combination with face recognition algorithm 2 The value of and in combination with face recognition algorithm 3 The value of can be different, and the same is true for face recognition algorithm 2 and face recognition algorithm 3.

[0055] Example 2

[0056] This embodiment takes 30,000 images as the detection set, and performs a performance test on the method of Embodiment 1 in a database of 1 billion images.

[0057] The parameters of the dynamic adaptive weighted face recognition multi-algorithm hybrid fusion in this embodiment are optimized as shown in Table 1:

[0058] Table 1

[0059]

[0060]

[0061] Experimental results: The recognition accuracy of face recognition algorithm 1 is 91.4%, the recognition accuracy of face recognition algorithm 2 is 90.6%, the recognition accuracy of face recognition algorithm 3 is 92.5%, and the recognition accuracy after fusion is 94.2%. Therefore, the method of Example 1 effectively improves the recognition accuracy of a single algorithm model.

[0062] For those skilled in the art, various corresponding changes and modifications can be made according to the above technical solutions and concepts, and all of these changes and modifications should be included in the protection scope of the claims of the present invention.

Claims

1. A dynamic adaptive weighted face recognition multi-algorithm hybrid fusion method, characterized in that: The steps include: S1. Normalization of evaluation system: Normalize each face recognition algorithm according to the following formula: Among them, x i is the similarity score obtained by comparing the i-th set of face recognition algorithms, T i is the similarity threshold set by the i-th face recognition algorithm, s i is the similarity score obtained by the i-th face recognition algorithm after normalization through the evaluation system; S2. Determine the adaptive weighting function: The total interval of similarity score [0,100] is divided into three intervals, [0,t1], (t 1, t2), [t2, 100], where the threshold T∈(t 1, t2), the adaptive weighting function is as follows: Among them, s i is the similarity score obtained by normalizing the evaluation system for the i-th face recognition algorithm, t1 and t2 are the interval values ​​corresponding to the i-th face recognition algorithm, r0 is the weight value corresponding to the similarity threshold T of the i-th face recognition algorithm, that is, the minimum value of the weight function, r2 is the weight value corresponding to t2, and r1 is the weight value corresponding to t1; S3, Hybrid Fusion: Through the two fusion methods of unary fusion and binary fusion, the similarity scores output by each face recognition algorithm are mixed and fused. The similarity score Output after mixed fusion is calculated as follows: In the above formula, A represents the result of the unary fusion strategy, and B represents the result of the binary fusion strategy. represents the weight of face recognition algorithm i in the unary fusion strategy, Indicates the weight of face recognition algorithm i in the binary fusion strategy. The same face recognition algorithm i is combined with different face recognition algorithms in the binary fusion strategy. The values ​​of are the same or different; and are the weights of the unary fusion strategy and the binary fusion strategy respectively; They represent the weights of the similarity scores output by face recognition algorithm i, calculated by formula (3); f(s a ,s b ) is the new model formed by combining face recognition algorithm a and face recognition algorithm b in the binary fusion strategy, a,b=1,2,…,N and a≠b, is f(s a ,s b ) is calculated by formula (3); a and b are non-repeating and disordered combinations of all face recognition algorithms participating in the hybrid fusion, and N represents the total number of face recognition algorithms participating in the hybrid fusion.

2. The method according to claim 1, characterized in that In step S3, there are:

3. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 2 is implemented.

4. A computer device, characterized in that: The method comprises a processor and a memory, wherein the memory is used to store a computer program; when the processor is used to execute the computer program, the method described in any one of claims 1 to 2 is implemented.

Images