A Method for Selecting Internal Feature Data of Object Recognition and Its Application

Abstract

The present invention provides a method for selecting internal feature data for object recognition and its application. The method for selecting internal feature data for object recognition includes the steps of: in response to a recognition request for an object to be recognized, extracting the feature data of the object to be recognized; using the feature data and a class label as parameters and inputting them into a selection model for internal feature data of object recognition to obtain the discrimination ability of the feature data; performing compression processing on the discrimination ability of the feature data to distinguish the feature data. Through the method for selecting internal feature data for object recognition and its application disclosed by the present invention, the recognition efficiency and accuracy can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of neural networks, and particularly to a method for selecting internal feature data for object recognition and its application. Background Art

[0002] A feature refers to an abstract result of the characteristics possessed by a research object. Features are the premise of pattern recognition. In our real life, a research object has many features. For example, in the process of face recognition, a large number of feature data are read from a human face. In fact, many of the feature data are irrelevant data, and there are not many feature data that can identify the object. In the process of pattern recognition, a large amount of irrelevant feature data will interfere with the calculation, resulting in inaccurate calculation and increasing the computing burden of the computer at the same time. Summary of the Invention

[0003] In view of the above-mentioned disadvantages of the prior art, the purpose of the present invention is to provide a method for selecting internal feature data for object recognition and its application. The present invention associates feature data with class labels, thereby screening the feature data to eliminate the feature data irrelevant to the class labels, so as to retain the feature data relevant to the class labels.

[0004] To achieve the above object and other related objects, the present invention provides a method for selecting internal feature data for object recognition, including:

[0005] In response to an identification request of an identification object, extracting the feature data of the identification object;

[0006] Taking the feature data and the class label as parameters and inputting them into a selection model of internal feature data for object recognition to obtain the discriminative ability of the feature data;

[0007] Performing compression processing on the discriminative ability of the feature data to distinguish the feature data.

[0008] In an embodiment of the present invention, screening is performed according to the discrimination result of the feature data, eliminating the feature data irrelevant to the class label, retaining the feature data relevant to the class label, and transmitting the feature data relevant to the class label to the hidden layer of the selection model of internal feature data for object recognition.

[0009] In an embodiment of the present invention, the step of taking the feature data and the class label as parameters and inputting them into a selection model of internal feature data for object recognition to obtain the discriminative ability of the feature data includes:

[0010] Taking the feature data and the class label as parameters and inputting them into a selection model of internal feature data for object recognition for zero-centralization processing to obtain the covariance correlation degree between the feature data and the class label;

[0011] Process the covariate correlation degree between the feature data and the class label to obtain the importance of the feature data relative to the class label;

[0012] Process the importance of the feature data relative to the class label to obtain the discriminative ability of the feature data.

[0013] In an embodiment of the present invention, the covariate correlation degree ρ between the feature data and the class label kj , is expressed as: where k is the k-th feature data,

[0014] j is the j-th class label,

[0015] n is the total number of training data,

[0016] x i is the i-th training data, x = [x1, x2,..., x n , x i ∈R p , R is the set of real numbers, p is the dimension of the feature,

[0017] y i is the i-th class label, y = [y1, y2,..., y n , y i ∈{0, 1} c , c is the total number of class labels.

[0018] In an embodiment of the present invention, the importance s of the feature data relative to the class label y , is expressed as: where

[0019] k is the k-th feature data,

[0020] y is the class label y,

[0021] j = 1, 2, 3,..., c.

[0022] In an embodiment of the present invention, the discriminative ability s of the feature data k , is expressed as:

[0023] In an embodiment of the present invention, the step of performing compression processing on the discriminative ability of the feature data to distinguish the feature data includes:

[0024] Process the discriminative ability of the feature data to obtain a weighted vector of corresponding weights;

[0025] Perform sparsity constraint processing on the weighted vector to obtain the selection result of the feature data;

[0026] Distinguish the feature data according to the selection result of the feature data.

[0027] In an embodiment of the present invention, the selection result of the feature data is expressed as: where

[0028] l(θ) is the loss function of the selection model of the feature data in object recognition,

[0029] d is the weighted vector corresponding to the weight, d=(d k )∈R p ,d k =1 - s k ,

[0030] λ1 and λ2 are weighting coefficients, 0≤λ1≤1, 0≤λ2≤1,

[0031] w is the weight coefficient vector connected between the input layer and the weighted input layer,

[0032] p and q are coefficients, 0≤p≤1, q>1,

[0033] K is the number of hidden layers in the selection model of the feature data in object recognition,

[0034] W (i) is the weight matrix connected between the (i - 1)-th layer network layer and the i-th layer network layer.

[0035] The present invention also provides a selection model of the feature data in object recognition, including:

[0036] An extraction module for extracting the feature data of the recognition object;

[0037] An operation module for calculating the feature data and the class label to obtain the discrimination ability of the feature data;

[0038] A compression module for compressing the discrimination ability of the feature data to distinguish the feature data; and

[0039] A screening module for screening the feature data, removing the feature data irrelevant to the class label, and retaining the feature data relevant to the class label.

[0040] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the steps of the method for selecting the feature data of object recognition are implemented.

[0041] As described above, the present invention provides a method for selecting object recognition feature data and its application, which can extract feature data related to class labels and eliminate feature data unrelated to class labels, thereby reducing the number of feature data and enhancing the correlation value between the feature data and the class labels, thus greatly improving the recognition efficiency and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for describing the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0043] Figure 1 It shows a flowchart of a method for selecting internal feature data for object recognition according to the present invention.

[0044] Figure 2 It shows a flowchart of sub-steps of step S2 in the method for selecting internal feature data for object recognition according to the present invention.

[0045] Figure 3 It shows a flowchart of sub-steps of step S3 in the method for selecting internal feature data for object recognition according to the present invention.

[0046] Figure 4 It shows a schematic structural diagram of a system for selecting internal feature data for object recognition according to the present invention.

[0047] Description of Component Labels:

[0048] 1. Extraction module; 2. Operation module; 3. Compression module; 4. Screening module. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0049] 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 a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0050] Please refer to Figure 1As shown in the figure, this embodiment provides a method for selecting internal feature data for object recognition. The method for selecting internal feature data for object recognition can be applied to one or more electronic devices. An electronic device is a device that can automatically perform numerical calculations and / or information processing according to pre-set or stored instructions. An electronic device can be any electronic product that can interact with a user, such as a personal computer, a tablet computer, etc. The hardware of an electronic device may include a processor, a memory, an application-specific integrated circuit, a programmable gate array, an embedded device, and a readable storage medium, etc. Among them, the processor and the memory can be coupled to each other. Program instructions can be stored in the memory. When the program instructions stored in the memory are executed by the processor, the method for selecting internal feature data for object recognition can be executed. Program instructions can also be stored in the readable storage medium. Thus, when the readable storage medium runs on the electronic device, the electronic device can execute the method for selecting internal feature data for object recognition. The method for selecting internal feature data for object recognition can be applied to object recognition processes such as face recognition and image recognition, and can be used to screen the feature data generated during the recognition process to reduce the subsequent calculation amount. The method for selecting internal feature data for object recognition may include:

[0051] Step S1: In response to a recognition request of a recognition object, extract the feature data of the recognition object.

[0052] In step S1, when recognizing a recognition object, a recognition system, etc. will scan the recognition object to obtain the feature data of the recognition object. The amount of information of the feature data of the recognition object is large, and it contains a large amount of feature data that is not relevant to the category label. The category label refers to a special attribute that an object has during the recognition process.

[0053] Please refer to Figure 2 As shown in the figure, step S2: Use the feature data and the category label as parameters and input them into the selection model of internal feature data for object recognition to obtain the discriminative ability of the feature data. The step of using the feature data and the category label as parameters and inputting them into the selection model of internal feature data for object recognition to obtain the discriminative ability of the feature data includes:

[0054] Step S21: Use the feature data and the category label as parameters and input them into the selection model of internal feature data for object recognition for zero-centering processing to obtain the covariation correlation degree ρ of the feature data and the category label kj , which is expressed as: Among them,

[0055] k is the kth feature data,

[0056] j is the jth category label,

[0057] n is the total number of training data,

[0058] x i is the i-th training data, x = [x1, x2,..., x n , x i ∈R p , where R is the set of real numbers and p is the dimension of the features.

[0059] y i is the i-th class label, y = [y1, y2,..., y n , y i ∈{0, 1} c , and c is the total number of class labels.

[0060] Step S22: Process the covariate correlation ρ kj between the feature data and the class label to obtain the importance of the feature data relative to the class label. Among them, the importance sk y of the k-th feature data relative to the class label y is expressed as:

[0061] Step S23: Process the importance of the feature data relative to the class label to obtain the discriminative ability of the feature data. Among them, the discriminative ability sk k of the k-th feature data is expressed as:

[0062] In step S21, the feature data refers to the relevant data obtained through recognition during the object recognition process. In order to correctly identify the feature data, the parameters in the recognition system need to be learned using the training data, so that the learned recognition network can be used for recognition. Therefore, it is recorded that the n training data in the recognition network have p-dimensional features, which can be expressed as x = [x1, x2,..., x n , x i ∈R p , and the class labels corresponding to the n training data can be expressed as y = [y1, y2,..., y n , y i ∈{0, 1} c , and thus the covariate correlation ρ kj between the k-th feature data and the j-th class label in the class label can be obtained.

[0063] In step S22, the higher the importance of the feature data relative to the class label, the higher the importance of the feature data, which indicates that the discriminative ability of the feature data is better. Since the importance sk y of the k-th feature vector relative to the class label y changes greatly, it cannot be directly judged, so it needs to be further processed to make the discriminative ability sk k of the k-th feature data within a certain numerical range.

[0064] In step S23, since the total number of class labels is c, for the class label y, first, according to the covariance relevance ρ kj between the k-th feature vector and the j-th class label vector, thus according to the importance s y of the k-th feature vector with respect to the class label y, the average value of the importance of the k-th feature vector with respect to all class labels is obtained, and thus the discriminative ability s k of the k-th feature data can be obtained. Since the discriminative ability s k of the k-th feature data ranges from 0 to 1, when the discriminative ability s k of the k-th feature data is closer to 1, it indicates that the discriminative ability of the corresponding feature data is stronger. When the discriminative ability s k of the k-th feature data is closer to 0, it indicates that the corresponding feature data has basically no discriminative ability. Therefore, when selecting feature data, try to retain the feature data with the discriminative ability s k of the k-th feature data closer to 1, and eliminate the feature data with the discriminative ability s k of the k-th feature data closer to 0.

[0065] In order to be able to judge the discriminative ability s k of the k-th feature data, a first threshold and a second threshold can be preset within the range of 0 to 1, where the first threshold can be greater than the second threshold. Thus, when the discriminative ability s k of the k-th feature data is greater than or equal to the first threshold, the corresponding feature data can be retained. When the discriminative ability s k of the k-th feature data is less than or equal to the second threshold, the corresponding feature vector can be eliminated. For the discriminative ability s k of the k-th feature data corresponding to the feature vector within the numerical range of the second threshold and the first threshold, it is impossible to directly judge it. Therefore, a sparsity constraint can be imposed on the discriminative ability s k of the k-th feature data, so that the feature data with the discriminative ability closer to 1 tends to 1 more quickly, and the feature data with the discriminative ability closer to 0 tends to 0 more quickly. In this embodiment, the magnitudes of the first threshold and the second threshold are not limited. For example, the first threshold can be 0.7, 0.8, 0.9, etc., and the second threshold can be 0.1, 0.2, 0.3, etc. The magnitudes of the first threshold and the second threshold can be set according to actual needs.

[0066] Please refer to Figure 3 as shown. Step S3: Perform compression processing on the discriminative ability of the feature data to distinguish the feature data. The steps of performing compression processing on the discriminative ability of the feature data to distinguish the feature data include:

[0067] Step S31: Process the discrimination ability of the feature data to obtain a weighted vector d with corresponding weights, expressed as: d = (d k ) ∈ R p , d k = 1 - s k .

[0068] Step S32: Perform sparsity constraint processing on the weighted vector d to obtain the selection result of the feature data expressed as: where

[0069] l(θ) is the loss function of the selection model of the feature data inside the object recognition,

[0070] λ1, λ2 are weighting coefficients, 0 ≤ λ1 ≤ 1, 0 ≤ λ2 ≤ 1

[0071] w is the weight coefficient vector connected between the input layer and the weighted input layer,

[0072] p, q are coefficients, 0 ≤ p ≤ 1, q > 1,

[0073] K is the number of hidden layers inside the selection model of the feature data in object recognition,

[0074] W (i) is the weight matrix connected between the (i - 1)-th layer network layer and the i-th layer network layer.

[0075] Step S33: According to the selection result of the feature data, distinguish the feature data.

[0076] In step S31, a weighted vector d can be defined first on the weight coefficient vector w between the input layer and the weighted input layer, d = (d k ) ∈ R p , d k = 1 - s k . Thus, when the feature data is highly correlated with the class label, that is, when the discrimination ability s k of the k-th feature data approaches 1, the weight d k of the corresponding weighted vector approaches 0. When the feature data is not relevant to the class label, that is, when the discrimination ability s k of the k-th feature data approaches 0, the weight d k of the corresponding weighted vector approaches 1.

[0077] In step S32, for a deep network model with K hidden layers, its model parameters can be expressed as: θ = {W (1) , b (1) , …, W (K+1) , b(K+1)}, where W (i) represents the weight matrix connecting the (i - 1)-th layer and the i-th layer, and the size of W (i) is denoted as n i-1 ×n i , and n i represents the number of nodes in the i-th layer, and b (i) is the bias vector of the i-th layer. For the selection model of internal feature data in object recognition, a weighted input layer is added on the deep network model. The weighted input layer is located between the input layer and the first hidden layer, so that the neurons in the input layer are correspondingly connected to the neurons in the weighted input layer, and any neuron in the weighted input layer is connected to any neuron in the first hidden layer. In this embodiment, the weight matrix between the input layer and the weighted input layer is denoted as W (1) , the bias vector is denoted as b (1) , and the weight coefficient vector between the input layer and the weighted input layer is denoted as w. Thus, the model parameters of the entire selection model of internal feature data in object recognition can be expressed as: θ = {w, W (1) , b (1) , …, W (K+1) , b (K+1)}. Then the corresponding loss function l(θ) can be expressed as: where

[0078] is the output feature of the K-th hidden layer corresponding to the input x j , x j is the j-th training data,

[0079] p is the probability density function,

[0080] y i is the i-th class label,

[0081] T is the vector transpose,

[0082] w l is the weight vector of the output layer corresponding to the i-th label class,

[0083] b l is the bias of the output layer corresponding to the i-th label class.

[0084] At this time, it is necessary to perform L1-norm sparsity constraint processing on the weight coefficient vector w between the input layer and the weighted input layer and the weight matrices W (i) of each subsequent hidden layer, and the depth feature selection function with L1-norm constraint based on the regression model can be obtained as: where the L1-norm refers to the sum of the absolute values of each element in the vector. When minimizing the above L1-norm sparsity constraint based on the regression model, a part of the weight coefficients w kis 0, which is equivalent to only selecting the weight coefficient w k The non-zero partial feature data participates in the subsequent network information transmission, and the weight coefficient w k The non-zero partial feature data represents the feature data associated with the class label, so as to achieve the purpose of deep feature selection. Although the above objective function with L1-norm sparsity constraint based on the regression model is non-convex and non-differentiable, the backpropagation in deep optimization does not require explicitly finding the explicit expression of the gradient to optimize. Therefore, some existing deep neural network toolkits can be used to optimize the objective function with L1-norm sparsity constraint based on the regression model.

[0085] In order to optimize the objective function with L1-norm sparsity constraint, in this embodiment, the idea that the L1-norm mixed with L2-norm constraint in the elastic net EN has more robust characteristics can be used to perform L2-norm constraint processing on the deep feature selection with regression model-based constraints. Thus, the robust deep feature selection function based on the elastic net EN constraint is expressed as: where the L2-norm refers to the sum of the squares of each element in the vector, and then take the square root.

[0086] Of course, the L1-norm sparsity constraint on the weight coefficient vector w of the weighted input layer in the deep feature selection with L1-norm sparsity constraint based on the regression model can be generalized to the general form of L p -norm (0 ≤ p ≤ 1) norm sparsity constraint, then the objective function of deep feature selection based on L p -norm constraint can be expressed as: where L p -norm refers to the sum of the p-th powers of each element in the vector, and then take the p-th root. In this embodiment, the general form of vector mixed norm can also be used to perform sparsity constraint on the weight coefficient vector w of the weighted input layer. By using the L p -norm (0 ≤ p ≤ 1) mixed with L q -norm (q > 1) method, the deep feature selection function based on vector mixed norm constraint is obtained, which is expressed as: where, L q -norm refers to the sum of the q-th powers of each element in the vector, and then take the q-th root.

[0087] In this embodiment, selecting the feature data that is highly correlated with the class label is more beneficial to the subsequent classification and recognition tasks. Therefore, it is necessary to make full use of the covariate correlation information between the features and the class label in the training data to assist in selecting discriminant features that are more beneficial to the subsequent classification and recognition. Therefore, the new features selected based on the regression model with maximized covariate correlation have stronger discriminant performance. For the weighted constraint of the regression model, d can be used as the weighted vector corresponding to the weight, which is expressed as: ||dn w|| 1. , thus the mixed norm constraint combining the elastic net EN weighted net constraint of the L1 norm and the L2 norm can be expressed as: ||d n w||1 + r||d n w|| 2. , then the mixed norm constraint combining the L p norm and the L q norm can be expressed as: ||d n w|| p + r||d n w|| q. .

[0088] In order to be able to select feature data that is highly relevant to the class label as much as possible, the weighted constraint of the regression model can be minimized, expressed as: where is the Hadamard product, that is, multiplying corresponding elements. When minimizing the weighted constraint of the above regression model, a larger-weight d k will impose a greater compression force on the corresponding weight coefficient w k , so that the weight coefficient w k tends to 0 more quickly, while a smaller-weight d k will not affect the compression of the corresponding weight coefficient w k , making it more inclined to retain the original value. Therefore, a weighted sparsity constraint of the regression model can be added to the weight coefficient vector w of the weighted input layer, and the objective function of the depth feature selection based on the weighted sparsity constraint of the regression model can be expressed as: Minimizing the objective function of the depth feature selection based on the weighted sparsity constraint of the regression model can make the weight coefficient w k corresponding to the feature data with low relevance to the class label be 0, so as to retain the weight coefficient w k corresponding to the feature data with high relevance to the class label be non-0, thus equivalent to selecting some feature data with high relevance to the class label to participate in the subsequent network information transmission, achieving the purpose of discriminative feature selection.

[0089] In this embodiment, in order to obtain a more general form of the objective function, a weighted sparsity constraint can be combined with a general form of the vector mixed norm to perform a sparsity constraint on the weight vector coefficient w of the weighted input layer. By the way of the L p norm (0 ≤ p ≤ 1) mixed with the L q norm (q > 1), the general objective function of the depth feature selection based on the vector mixed norm constraint is obtained, expressed as: In this general objective function, when p = 1 and r = 0, the general objective function is transformed into a depth discriminant feature selection function based on the weighted sparse constraint of the regression model; when p = 1 and q = 2, the general objective function is transformed into a depth discriminant feature selection function based on the weighted sparse constraint of the elastic net EN; when r = 0, the general objective function is transformed into a depth discriminant feature selection function based on the weighted sparse constraint of the L p -norm weighted sparse constraint.

[0090] In step S33, the general objective function of the depth feature selection based on the vector mixed norm constraint can be minimized, so as to obtain the selection result of the feature data. At this time, for the feature data, some of the feature data are associated with the class label, while some of the feature data are not related to the class label.

[0091] Step S4: Screen according to the discrimination result of the feature data, eliminate the feature data that is not related to the class label, retain the feature data that is related to the class label, and transfer the feature data that is related to the class label to the hidden layer of the feature data selection model in the object recognition.

[0092] In step S4, according to the selection result of the feature data, the feature data can be divided into feature data that is not related to the class label and feature data that is related to the class label. In order to reduce the computational complexity, the feature data that is not related to the class label can be eliminated, so as to only retain the feature data that is related to the class label. At this time, the number of feature data is greatly reduced. At the same time, the feature data that is related to the class label is transferred to the hidden layer of the feature data selection model in the object recognition for object recognition, so as to reduce the computing burden of the computer and the calculation result is accurate.

[0093] Please refer to Figure 4 As shown, this embodiment also discloses a feature data selection system in object recognition,

[0094] The selection system for internal feature data of object recognition may include an extraction module 1, an operation module 2, a compression module 3, and a screening module 4. Among them, the extraction module 1 can be used to extract the feature data of the recognition object. The operation module 2 can use the extracted feature data and the class label as parameters to obtain the discrimination ability of the corresponding feature data. The compression module 3 can be used to perform compression processing on the discrimination ability of the feature data, making the weight coefficient corresponding to the feature data with low relevance to the class label be 0, so as to retain the weight coefficient corresponding to the feature data with high relevance to the class label as non-0, thereby distinguishing the feature data. The screening module 4 performs screening according to the discrimination result of the feature data, eliminates the feature data irrelevant to the class label, retains the feature data relevant to the class label, and transmits the feature data relevant to the class label to the hidden layer of the selection model of the internal feature data of object recognition.

[0095] In summary, through a method for selecting feature data of object recognition and its application provided by the present invention, it is possible to extract feature data related to the class label and eliminate feature data irrelevant to the class label, thereby reducing the number of feature data and enhancing the correlation value between the feature data and the class label, thus greatly improving the recognition efficiency and accuracy.

[0096] In the description of this specification, the descriptions referring to terms such as "this embodiment", "example", "specific example", etc. mean that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in a suitable manner in any one or more embodiments or examples.

[0097] The embodiments of the present invention disclosed above are only used to help explain the present invention. The embodiments do not describe all details in detail, nor limit the invention to the specific embodiments described. Obviously, many modifications and variations can be made according to the content of this specification. This specification selects and specifically describes these embodiments to better explain the principle and practical application of the present invention, so that those skilled in the art can understand and utilize the present invention well. The present invention is only limited by the claims and their full scope and equivalents.

Claims

1. A method for selecting internal feature data of object recognition, characterized in that, Applied to the field of image recognition, including: In response to an identification request of an identification object, extracting feature data of the identification object; Taking the feature data and a class label as parameters, inputting them into a selection model of in-object recognition feature data for zero-centering processing to obtain the covariate correlation between the feature data and the class label; Processing the covariate correlation between the feature data and the class label to obtain the importance of the feature data relative to the class label; Processing the importance of the feature data relative to the class label to obtain the discriminative ability of the feature data; Performing compression processing on the discriminative ability of the feature data to distinguish the feature data; Among them, the covariance correlation degree between the feature data and the class label is expressed as: where is the th feature data, is the th class label, is the total number of training data, is the th training data, , is the set of real numbers, is the dimension of the feature, is the th class label, , is the total number of class labels; The importance of the feature data relative to the class label , expressed as: , where is the th feature data, is the class label , ; The discriminability of the feature data , expressed as: , .

2. The method for selecting internal feature data for object recognition according to claim 1, wherein Also including: According to the discrimination result of the feature data, screening, eliminating the feature data irrelevant to the class label, retaining the feature data relevant to the class label, and transmitting the feature data relevant to the class label to a hidden layer of a selection model of in-object recognition feature data.

3. The method for selecting internal feature data for object recognition according to claim 1, wherein The step of performing compression processing on the discriminative ability of the feature data to distinguish the feature data includes: Processing the discriminative ability of the feature data to obtain a weighted vector of corresponding weights; Performing sparsity constraint processing on the weighted vector to obtain a selection result of the feature data; According to the selection result of the feature data, distinguishing the feature data.

4. The method for selecting internal feature data for object recognition according to claim 3, wherein The selection result of the characteristic data , is expressed as: , where is the loss function of the selection model for the internal feature data in object recognition, , is the weighting coefficient, , , is a weighted vector for the corresponding weight, ; is the weight coefficient vector connected between the input layer and the weighted input layer. , is a coefficient, , is the number of hidden layers in the selection model for the internal feature data in object recognition, Is the weight matrix connected between the layer network layer and the layer network layer.

5. An object recognition internal feature data selection system, characterized in that, Applied to the field of image recognition, including: An extraction module for extracting feature data of an identification object; An operation module for taking the feature data and a class label as parameters, inputting them into a selection model of in-object recognition feature data for zero-centering processing to obtain the covariate correlation between the feature data and the class label; processing the covariate correlation between the feature data and the class label to obtain the importance of the feature data relative to the class label; processing the importance of the feature data relative to the class label to obtain the discriminative ability of the feature data; A compression module for performing compression processing on the discriminative ability of feature data to distinguish the feature data; and A screening module for screening the feature data, eliminating the feature data irrelevant to the class label, and retaining the feature data relevant to the class label; Among them, the correlation degree of the feature data with the covariate of the class label is expressed as: where is the th feature data, is the th class label, is the total number of training data, is the th training data, , is the set of real numbers, is the dimension of the feature, is the th class label, , is the total number of class labels; The importance of the feature data relative to the class label , expressed as: , where is the th feature data, is the class label , ; Discrimination ability of the feature data , expressed as: , .

6. An electronic device, characterized in that, Comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that when the processor executes the computer program, the steps of the method according to any one of claims 1 to 4 are implemented.

Images