Feature self-decomposition face recognition method and system with illumination robustness

Abstract

This invention discloses a feature-self-decomposition face recognition method and system with illumination robustness. Based on the grayscale values ​​of each face image, corresponding illumination labels are obtained to create a dataset. The network structure includes a backbone feature extraction network and a feature decomposition network. The backbone network extracts face image features, while the feature decomposition network decomposes these features into illumination-dependent and illumination-independent features. Illumination-independent features are the more discriminative face features used for subsequent face recognition. This invention addresses the face recognition problem under varying illumination conditions by using a feature decomposition network to decompose the extracted face image features into illumination-dependent and illumination-independent feature vectors, reducing the impact of illumination information on face feature vectors and improving the accuracy of face recognition.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of face recognition, in particular to a feature self-decomposition face recognition method and system with illumination robustness. BACKGROUND

[0002] Face pictures taken in real scenes are obviously affected by illumination. For example, pictures taken under the sun have a larger brightness, and pictures taken at night have a smaller brightness, which can easily cause dramatic changes in face features. According to existing face recognition technology, as long as a large number of face pictures with different exposure are provided, a neural network can learn face features with illumination robustness. This type of method mainly tries to weaken the influence of illumination by letting the neural network extract features of eyebrows, eyes, nose, mouth and other parts of the face, but in fact, this type of method does not process face images with illumination changes, so the illumination information is still embedded in the extracted face feature vector to some extent. This part of information can cause the difference between face images of the same identity to become larger and the difference between face images of different identities to become smaller, thereby affecting the accuracy of face recognition. SUMMARY

[0003] To solve the above problems, the present application provides a feature self-decomposition face recognition method and system with illumination robustness. The system uses a feature decomposition network to decompose the extracted face image features into illumination-dependent feature vectors and illumination-independent feature vectors to solve the problem of face recognition under illumination and improve the accuracy of face recognition.

[0004] In a first aspect of the present application, a feature self-decomposition face recognition method with illumination robustness includes the following steps:

[0005] Image preprocessing: for a given face data set, first perform face detection, align the detected face, and then scale the aligned face picture to a specified size;

[0006] Image label making: calculate the gray value of the preprocessed face picture, perform gray value normalization, generate the corresponding illumination label according to the normalized gray value, and obtain an image data set with illumination labels;

[0007] Constructing a feature self-decomposition face recognition network model: the feature self-decomposition face recognition network model includes a backbone feature extraction network and a feature decomposition network. The backbone feature extraction network is used to extract face image features, and the feature decomposition network is used to decompose the extracted face image features into illumination-dependent feature vectors and illumination-independent feature vectors;

[0008] Network model training: the network model is trained using an image dataset with illumination labels, the first stage trains the backbone feature extraction network, the second stage freezes the parameters of the trained backbone feature extraction network, and the feature decomposition network is trained;

[0009] Identification process: using the trained network model for identification, the identification process includes face verification and face recognition, wherein the face verification specifically includes: given two face pictures, through image preprocessing, the two face pictures are respectively sent into the feature decomposition face recognition network model, two illumination-independent feature vectors are obtained, the similarity between the two illumination-independent feature vectors is calculated, when the similarity is greater than the set threshold, it is determined that the two face pictures have the same identity information, otherwise, the two face pictures come from different people; face recognition specifically includes: storing a large number of face feature vectors in the feature library, the face feature vectors are all illumination-independent feature vectors extracted by the trained feature decomposition face recognition network model, when a given face picture is input into the feature decomposition face recognition network model to obtain a feature vector, the face picture with the maximum similarity value is found in the feature library according to the feature vector, and the identity information corresponding to the face picture is determined according to the face picture with the maximum similarity value, if the similarity value is greater than the set threshold, the identity information of the corresponding face in the feature library is output, otherwise, it is determined as a stranger face.

[0010] Further, the image preprocessing specifically includes: using the MTCNN method for face detection, detecting the face position area while obtaining the position information of the five key points of the face, using the position information of the five key points and the standard position relationship to obtain the affine transformation matrix, and aligning the face through the affine transformation matrix.

[0011] Further, the image label making specifically includes: unifying the gray value of each face picture in the [0,1] interval, v max is the maximum value of the image gray value, v min is the minimum value of the image gray value, wherein the label value of v max is 1, the label value of v min is 0, and the label value of other image gray values v varies in the [0,1] interval according to the obtained gray value size, and the label value p is calculated as follows:

[0012]

[0013] Further, the backbone feature extraction network adopts light-FaceNet, which contains four bottlenecks in total, each bottleneck contains a dimension reduction operation and a dimension increase operation, and the backbone feature extraction network finally generates a 512-dimensional feature vector.

[0014] Further, the feature decomposition network receives the n-dimensional feature vector V extracted by the backbone feature extraction network, generates a weight vector W through a weight generation module in the feature decomposition network, multiplies the generated weight vector W with the n-dimensional feature vector V to obtain an illumination-related feature vector B, obtains a vector 1-W according to the weight vector W, and multiplies the vector 1-W with the n-dimensional feature vector V to obtain an illumination-independent feature vector F after removing the illumination information.

[0015] Further, the weight generation module is composed of two fully connected layers, an activation function, a batch normalization layer and a sigmoid layer, wherein one of the two fully connected layers is used to reduce the dimension of the feature vector V, the other is used to increase the dimension of the reduced feature vector V, and the sigmoid layer is used to limit the element value of the corresponding position in the weight vector W to be between [0, 1].

[0016] Further, two loss functions are designed in the network model training process for supervised learning, wherein the smooth-L1 loss is used for regression of the illumination-related feature vector, and the smooth-L1 loss is mainly calculated by regression of the predicted gray value and the gray value generated by the label; the ArcFace angle loss is used for regression of the face image feature loss, and the calculation formulas of the smooth-L1 loss and the ArcFace angle loss are as follows:

[0017]

[0018]

[0019] Wherein, x represents the difference between the label value and the predicted value, N represents the number of batch samples, θ is the angle interval between the sample and the classification boundary, y i is the class label, j is the class boundary mark, i is the current face class, m is the angle interval, and S is the radius.

[0020] In the second aspect of the present application, a feature decomposition face recognition system with illumination robustness is provided, comprising:

[0021] An image preprocessing unit is used for face detection on a given face data set, face alignment on the detected face, and scaling of the aligned face picture to a specified size.

[0022] An image label making unit is used for calculating the gray value of the preprocessed face picture, performing gray value normalization, generating the corresponding illumination label according to the normalized gray value, and obtaining the image data set with the illumination label.

[0023] The unit for constructing a feature self-decomposition face recognition network model is configured to construct a feature self-decomposition face recognition network model, which comprises a backbone feature extraction network and a feature decomposition network.

[0024] The network model training unit is configured to train the network model by using an image data set with illumination labels, train the backbone feature extraction network in a first stage, freeze the parameters of the trained backbone feature extraction network in a second stage, and train the feature decomposition network.

[0025] The recognition unit is configured to perform recognition by using the trained network model, and the recognition process comprises face verification and face recognition.

[0026] In a third aspect, the present application provides a feature self-decomposition face recognition system with illumination robustness, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the feature self-decomposition face recognition method with illumination robustness.

[0027] In a fourth aspect, the present application provides a computer-readable storage medium having instructions stored thereon, wherein the instructions, when executed by a processor, cause the processor to perform the feature self-decomposition face recognition method with illumination robustness.

[0028] This invention provides a lighting-robust feature-self-decomposition face recognition method and system. Based on the grayscale values ​​of each face image, corresponding lighting labels are obtained to create a dataset. The network structure consists of two parts: a backbone feature extraction network (light-FaceNet) and a feature decomposition network. The backbone network primarily extracts face image features, while the feature decomposition network decomposes these features into lighting-dependent and lighting-independent features. The lighting-independent features are the more discriminative face features used for subsequent face recognition verification. The network training process is divided into two stages. The first stage trains the backbone network, which is then used to extract robust face features. The second stage trains the feature decomposition network, decomposing the features extracted in the first stage. The main model optimization learns to find the feature elements in the feature vector obtained in the first stage that are more lighting-dependent. This invention utilizes an eigenvalue decomposition network to decompose extracted facial image features into illumination-dependent feature vectors and illumination-independent feature vectors, thereby solving the problem of face recognition under illumination conditions, reducing the influence of illumination information on facial feature vectors, and improving the accuracy of face recognition. Attached Figure Description

[0029] Figure 1 This is a flowchart of a feature self-decomposition face recognition method with illumination robustness in an embodiment of the present invention;

[0030] Figure 2 This is a schematic diagram of the feature decomposition network method in an embodiment of the present invention;

[0031] Figure 3 This is a schematic diagram of the weight generation module method flow in an embodiment of the present invention;

[0032] Figure 4 This is a schematic diagram of the training process of the feature self-decomposition face recognition network model in an embodiment of the present invention;

[0033] Figure 5 This is a schematic diagram of the face verification process in an embodiment of the present invention;

[0034] Figure 6 This is a schematic diagram of the face recognition process in an embodiment of the present invention;

[0035] Figure 7 This is a schematic diagram of the structure of a feature self-decomposition face recognition system with illumination robustness in an embodiment of the present invention;

[0036] Figure 8 This is an architectural diagram of the computer device in an embodiment of the present invention. Detailed Implementation

[0037] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the drawings, not the entire structure.

[0038] Before discussing the exemplary embodiments in more detail, it should be noted that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the steps as sequential processes, many of these steps can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the steps can be rearranged. The process can be terminated when its operation is complete, but may also have additional steps not included in the figures. The process can correspond to a method, function, procedure, subroutine, subroutine, etc.

[0039] This invention provides a method and system for self-decomposition of features with illumination robustness, as illustrated in the following embodiments:

[0040] Based on Embodiment 1 of the present invention

[0041] like Figure 1 The diagram shown is a flowchart of a feature self-decomposition face recognition method with illumination robustness according to Embodiment 1 of the present invention. The specific steps are as follows:

[0042] S110. Image preprocessing: For a given face dataset, first perform face detection, then perform face alignment on the detected faces, and finally scale the aligned face images to the specified size.

[0043] In a preferred embodiment, image preprocessing specifically includes: using the MTCNN method to perform face detection, simultaneously obtaining the position information of five key points of the face when the face location region is detected, using the position information of the five key points and the standard positional relationship to obtain the affine transformation matrix, and aligning the face using the affine transformation matrix.

[0044] Specifically, for a given face dataset, face detection is first performed, followed by face alignment. The aligned face images are then scaled to a specified size, preferably 112*112. For face detection, the MTCNN method is used to detect the face location region and simultaneously acquire the positional information of five key facial points. Preferably, these five key points are the center points of the left and right eyes, the tip of the nose, and the corners of the mouth. Using the positions of these five key points and their standard positional relationships, an affine transformation matrix is ​​calculated, and this matrix is ​​used to align the faces.

[0045] S120. Image labeling: Calculate the grayscale value of the preprocessed face image and normalize the grayscale value to generate the corresponding illumination label, thereby obtaining an image dataset with illumination labels.

[0046] In a preferred embodiment, image tagging specifically includes: unifying the grayscale values ​​of each face image within the range of [0,1], v max v represents the maximum grayscale value of the image. min The minimum value of the image grayscale, where v max The label value is 1, v min The label value of the grayscale value v is 0. The label values ​​of other image grayscale values ​​v vary within the range of [0,1] depending on the obtained grayscale value. The formula for calculating the label value p is as follows:

[0047]

[0048] Specifically, based on the grayscale value of each face image, a corresponding illumination label is obtained to create a dataset. Existing face datasets lack illumination-related labels. Therefore, the inventors calculated the grayscale values ​​of all preprocessed face images and normalized them, unifying the grayscale value of each face image within the [0,1] interval. Specifically, the normalized grayscale value is the label value. The formula for calculating the label value p is shown above, v max v represents the maximum grayscale value of the image. min The minimum value of the image grayscale, where v max The label is 1, v min The label for the first image is 0, and the label value p for the grayscale values ​​of other images varies within the range [0,1] depending on the obtained grayscale value. The greater the brightness, the closer the label value is to 1, and the smaller the brightness, the closer the label value is to 0.

[0049] S130. Construct a feature-self-decomposition face recognition network model: The feature-self-decomposition face recognition network model includes a backbone feature extraction network and a feature decomposition network. The backbone feature extraction network is used to extract face image features, and the feature decomposition network is used to decompose the extracted face image features into illumination-dependent feature vectors and illumination-independent feature vectors. Among them, the illumination-independent feature vectors are more discriminative face features, which are used for subsequent face recognition verification.

[0050] In a preferred embodiment, the backbone feature extraction network uses light-FaceNet, which contains four bottlenecks. Each bottleneck contains a dimensionality reduction operation and a dimensionality increase operation. The backbone feature extraction network finally generates a 512-dimensional feature vector.

[0051] Specifically, the backbone network light-FaceNet structure is shown in Table 1 below. It is used to extract face image features. light-FaceNet contains four bottlenecks. Each bottleneck contains a dimensionality reduction operation and a dimensionality increase operation. The purpose is to further compress the number of parameters of the 3*3 convolution and finally generate a 512-dimensional feature vector.

[0052] Table 1 Main structure of light-FaceNet

[0053]

[0054] In a preferred embodiment, such as Figure 2 As shown, after the eigenvalue decomposition network receives the n-dimensional feature vector V extracted by the backbone feature extraction network, it generates a weight vector W through the weight generation module in the eigenvalue decomposition network. The generated weight vector W is multiplied with the n-dimensional feature vector V to obtain the illumination-related feature vector B. Based on the weight vector W, the vector 1-W is obtained. The vector 1-W is multiplied with the n-dimensional feature vector V to obtain the illumination-independent feature vector F after removing illumination information. The illumination-independent feature vector F is more discriminative.

[0055] In a preferred embodiment, the core of the feature decomposition network is the weight generation module, such as... Figure 3 As shown, the weight generation module consists of two fully connected layers, an activation function, a batch normalization layer, and a sigmoid layer. One of the fully connected layers reduces the dimensionality of the feature vector V, while the other increases the dimensionality of the reduced feature vector V. The sigmoid layer restricts the output values ​​of corresponding elements in the weight vector W to the range [0,1]. The weight generation module primarily learns which elements in the feature vector V are more relevant to illumination features and which are more relevant to facial features, and decomposes the illumination features from the original features to obtain more discriminative facial features independent of illumination. If a certain element in the feature vector V is mainly constructed from illumination features, then the corresponding element value in the output weight vector W will be closer to 1, and this information will be allocated more to the illumination features.

[0056] S140. Network Model Training: The network model is trained using an image dataset with illumination labels. The first stage trains the backbone feature extraction network, and the second stage freezes the parameters of the trained backbone feature extraction network and focuses on training the feature decomposition network.

[0057] In a preferred embodiment, two loss functions are designed for supervised learning during network model training. The illumination-related feature vectors are regressed using smooth-L1 loss, which is calculated primarily by regressing the predicted grayscale values ​​against the grayscale values ​​generated from the labels. The face image feature loss is regressed using ArcFace angle loss. The calculation formulas for smooth-L1 loss and ArcFace angle loss are as follows:

[0058]

[0059]

[0060] Where x represents the difference between the label value and the predicted value, N represents the number of samples in the batch, θ is the angular distance between the sample and the classification boundary, and y i is the category label, j is the category boundary marker, i is the current face category, m is the angular interval, and S is the radius.

[0061] Specifically, the generated illumination-related feature vector B is used to generate an element value q through a weight vector X, calculated as follows:

[0062]

[0063] The new feature element q is the predicted value obtained by multiplying the corresponding elements of the illumination feature vector B and the weight vector X, and then summing them. n represents the total number of elements. The predicted value q and the true value are then input into the smooth-L1 loss as the network's optimization objective loss, i.e., the optimization objective loss between the predicted grayscale value and the grayscale value generated by the label, used for model feedback training. The weight X is a learnable weight vector, similar to the parameters of the deep network layers, obtained through loss value backpropagation and stochastic gradient descent algorithm, and continuously updated.

[0064] Furthermore, the training process is as follows: Figure 4 As shown, the aligned and scaled face image is fed into the feature self-decomposition face recognition network model. First, face horizontal flipping data augmentation is performed with a probability of 0.5. The training process is mainly divided into two stages. The first stage trains the backbone network, light-FaceNet. This part is trained by directly adding ArcFace angle loss after the feature vector V, i.e., training a feature extractor. After this stage of training, a relatively robust face feature representation can be obtained. The second stage mainly trains the feature decomposition network. The parameters of this part of the network are mainly concentrated in the feature decomposition module. In this stage of training, the parameters of the backbone network are frozen, and the training focuses on the parameters of the feature decomposition module, mainly learning how to better decompose the feature vector V into illumination-dependent feature vectors and illumination-independent feature vectors.

[0065] In a preferred embodiment, the stochastic gradient descent algorithm is used for training, with momentum set to 0.9 and weight decay parameter set to 0.0001. In face recognition training, to ensure more stable training and convergence to the optimal effect, a larger batch size is required. In the preferred embodiment, this is empirically set to 512, with an initial learning rate of 0.1, which is reduced to 0.1 times the original rate in the 60th and 80th rounds respectively. The training in both stages terminates at the 100th round.

[0066] S150. Recognition Process: Recognition is performed using a trained network model. The recognition process includes face verification and face recognition. Face verification specifically includes: given two face images, image preprocessing is performed, and the two face images are fed into a feature self-decomposition face recognition network model to obtain two illumination-independent feature vectors. The similarity between the two illumination-independent feature vectors is calculated. When the similarity is greater than a set threshold, the two face images are determined to have the same identity information; otherwise, the two face images are from different people. Face recognition specifically includes: storing a large number of face feature vectors in a feature library. The face feature vectors are all illumination-independent feature vectors extracted using the trained feature self-decomposition face recognition network model. When a face image is given, the feature vector is obtained by inputting it into the feature self-decomposition face recognition network model. Based on the feature vector, the face image with the highest similarity value is found in the feature library, and its corresponding identity information is determined. If the similarity value is greater than a set threshold, the identity information of the corresponding face in the feature library is output; otherwise, it is determined to be a stranger's face.

[0067] In a preferred embodiment, the face verification process is as follows: Figure 5 As shown, given two face images, image preprocessing is performed, and the two face images are fed into a feature self-decomposition face recognition network model to obtain attitude-independent feature vectors. The similarity between the two feature vectors is calculated, and cosine similarity is used. When the similarity is greater than the set threshold, where the threshold is preferably 0.4, it is determined that the two face images have the same identity information; otherwise, the two images come from different people.

[0068] In a preferred embodiment, the face recognition process is as follows: Figure 6 As shown, a large number of facial feature vectors are stored in the feature library. These feature vectors are obtained by using a trained feature self-decomposition face recognition network model to extract illumination-independent feature vectors. Given a face image, the feature vectors are output. Based on the feature vectors, the face image with the highest similarity value is found in the feature library, and its corresponding identity information is determined. If the similarity value is greater than a set threshold, preferably 0.4, the information of that face is output; otherwise, it is determined to be a stranger.

[0069] Based on Embodiment 2 of the present invention

[0070] The illumination-robust feature self-decomposition face recognition system 700 provided in Embodiment 2 of the present invention can execute the illumination-robust feature self-decomposition face recognition method provided in any embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the execution method. The system can be implemented by software and / or hardware (integrated circuit) and can generally be integrated into a server or terminal device.

[0071] Figure 7 This is a schematic diagram of the structure of a feature self-decomposition face recognition system 700 with illumination robustness according to Embodiment 2 of the present invention. (Refer to...) Figure 7 The illumination-robust feature self-decomposition face recognition system 700 of this invention specifically includes:

[0072] The image preprocessing unit 710 is used to perform face detection on a given face dataset, perform face alignment on the detected faces, and then scale the aligned face image to a specified size.

[0073] The image labeling unit 720 is used to calculate the grayscale value of the preprocessed face image and normalize the grayscale value. Based on the normalized grayscale value, the corresponding illumination label is generated, and an image dataset with illumination label can be obtained.

[0074] A feature self-decomposition face recognition network model unit 730 is used to construct a feature self-decomposition face recognition network model. The feature self-decomposition face recognition network model includes a backbone feature extraction network and a feature decomposition network. The backbone feature extraction network is used to extract face image features, and the feature decomposition network is used to decompose the extracted face image features into illumination-dependent feature vectors and illumination-independent feature vectors.

[0075] The network model training unit 740 is used to train the network model using an image dataset with illumination labels. The first stage trains the backbone feature extraction network, and the second stage freezes the parameters of the trained backbone feature extraction network and focuses on training the feature decomposition network.

[0076] The recognition unit 750 is used for recognition using a trained network model. The recognition process includes face verification and face recognition. Face verification specifically includes: given two face images, image preprocessing is performed, and the two face images are fed into a feature self-decomposition face recognition network model to obtain two illumination-independent feature vectors. The similarity between the two illumination-independent feature vectors is calculated. When the similarity is greater than a set threshold, the two face images are determined to have the same identity information; otherwise, the two face images are from different people. Face recognition specifically includes: storing a large number of face feature vectors in a feature library. The face feature vectors are all illumination-independent feature vectors extracted using the trained feature self-decomposition face recognition network model. When a face image is given, the feature vector is obtained by inputting it into the feature self-decomposition face recognition network model. Based on the feature vector, the face image with the highest similarity value is found in the feature library, and its corresponding identity information is determined. If the similarity value is greater than a set threshold, the identity information of the corresponding face in the feature library is output; otherwise, it is determined to be a stranger's face.

[0077] In addition to the five units mentioned above, the system 700 may also include other components; however, since these components are not relevant to the embodiments of this disclosure, their illustrations and descriptions are omitted here.

[0078] The specific working process of a feature self-decomposition face recognition system 700 with illumination robustness is described in the above-described embodiment 1 of the feature self-decomposition face recognition method with illumination robustness, and will not be repeated here.

[0079] Based on Embodiment 3 of the present invention

[0080] The system according to embodiments of the present invention can also be used by means of Figure 8 The architecture of the computing device shown is used to implement this. Figure 8 The architecture of the computing device is shown. For example... Figure 8 As shown, the computer system 810 includes a system bus 830, one or more CPUs 840, input / output components 820, and a memory 850. The memory 850 can store various data or files used by the computer for processing and / or communication, as well as program instructions executed by the CPU, including the method of Embodiment 1. Figure 8 The architecture shown is merely exemplary and should be adjusted according to actual needs when implementing different devices. Figure 8 One or more components in.

[0081] Based on Embodiment 4 of the present invention

[0082] The embodiments of the present invention can also be implemented as a computer-readable storage medium. The computer-readable storage medium according to Embodiment 4 stores computer-readable instructions. When the computer-readable instructions are executed by a processor, the illumination-robust feature self-decomposition face recognition method according to Embodiment 1 of the present invention, described with reference to the above figures, can be performed.

[0083] Based on the above embodiments, the feature-self-decomposition face recognition method and system with illumination robustness are described. Based on the grayscale value of each face image, corresponding illumination labels are obtained to create a dataset. A network structure is designed, consisting of two parts: a backbone feature extraction network (light-FaceNet) and a feature decomposition network. The backbone network is mainly used to extract face image features, while the feature decomposition network decomposes the extracted face features into illumination-dependent features and illumination-independent features. Illumination-independent features are the more discriminative face features used for subsequent face recognition verification. The network training process is divided into two stages. The first stage trains the backbone network, which can then be used to extract more robust face features. The second stage mainly trains the feature decomposition network, decomposing the features extracted in the first stage. The main model optimization learning finds the feature elements in the feature vector obtained in the first stage that are more related to illumination. This invention utilizes an eigenvalue decomposition network to decompose extracted facial image features into illumination-dependent feature vectors and illumination-independent feature vectors, thereby solving the problem of face recognition under illumination conditions, reducing the influence of illumination information on facial feature vectors, and improving the accuracy of face recognition.

[0084] Note that the above description is merely a preferred embodiment of the present invention and the technical principles employed. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions can be made without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the above embodiments, and may include many other equivalent embodiments without departing from the concept of the present invention, the scope of which is determined by the scope of the appended claims.

Claims

1. A feature self-decomposition face recognition method with illumination robustness, characterized in that, The method comprises the following steps: Image preprocessing: for a given face data set, first perform face detection, then perform face alignment on the detected face, and then scale the aligned face picture to a specified size; Image label making: calculate the gray value of the preprocessed face picture, perform gray value normalization, generate the corresponding illumination label according to the normalized gray value, and obtain the image data set with the illumination label; Constructing a feature self-decomposition face recognition network model: the feature self-decomposition face recognition network model comprises a backbone feature extraction network and a feature decomposition network, the backbone feature extraction network is used to extract face image features, and the feature decomposition network is used to decompose the extracted face image features into an illumination-related feature vector and an illumination-independent feature vector; Network model training: the image data set with the illumination label is used for network model training, the first stage trains the backbone feature extraction network, the second stage freezes the parameters of the trained backbone feature extraction network, and the feature decomposition network is trained; The recognition process comprises face verification and face recognition, wherein the face verification specifically comprises: given two face pictures, the two face pictures are respectively input into the feature self-decomposition face recognition network model through image preprocessing, two illumination-independent feature vectors are obtained, the similarity between the two illumination-independent feature vectors is calculated, when the similarity is greater than a set threshold, it is determined that the two face pictures have the same identity information, otherwise, the two face pictures come from different people; the face recognition specifically comprises: a large number of face feature vectors are stored in a feature library, the face feature vectors are illumination-independent feature vectors extracted by the trained feature self-decomposition face recognition network model, when a given face picture is input into the feature self-decomposition face recognition network model to obtain a feature vector, the face picture with the maximum similarity value is found in the feature library according to the feature vector, and the identity information corresponding to the face picture is determined according to the face picture with the maximum similarity; if the similarity value is greater than a set threshold, the identity information of the corresponding face in the feature library is output, otherwise, it is determined as a stranger face; The feature decomposition network receives the n-dimensional feature vector V extracted by the backbone feature extraction network, generates a weight vector W through a weight generation module in the feature decomposition network, multiplies the generated weight vector W and the n-dimensional feature vector V to obtain an illumination-related feature vector B, obtains a vector 1-W according to the weight vector W, and multiplies the vector 1-W and the n-dimensional feature vector V to obtain an illumination-independent feature vector F after removing the illumination information; The weight generation module comprises two fully connected layers, an activation function, a batch normalization layer and a sigmoid layer, wherein one of the two fully connected layers is used to reduce the dimension of the feature vector V, the other is used to increase the dimension of the reduced feature vector V, and the sigmoid layer is used to limit the element value of the corresponding position in the weight vector W to [0, 1].

2. The feature self-dissolving face recognition method with illumination robustness according to claim 1, characterized in that, The image preprocessing specifically comprises: adopting an MTCNN method to perform face detection, acquiring position information of five key points of a face while detecting a face position area, using the position information of the five key points and a standard position relationship to obtain an affine transformation matrix, and aligning the face through the affine transformation matrix.

3. The feature self-dissolving face recognition method with illumination robustness according to claim 1, characterized in that, The image label making step specifically includes: uniformly setting the gray value of each face picture in the interval [0, 1], v max is the maximum value of the image gray value, v min is the minimum value of the image gray value, wherein v max The label value of v min is 0, and the label value of other picture gray values v varies in the interval [0, 1] according to the size of the obtained gray value, and the label value p is calculated according to the following formula:

4. The feature self-dissolving face recognition method with illumination robustness according to claim 1, characterized in that, The backbone feature extraction network adopts light-FaceNet, and the light-FaceNet comprises four bottlenecks in total, each bottleneck comprising a dimension reduction operation and a dimension increase operation, and the backbone feature extraction network finally generates a 512-dimensional feature vector.

5. The feature self-dissolving face recognition method with illumination robustness according to claim 1, characterized in that, Two loss functions are designed for supervised learning in the network model training process, wherein a smooth-L1 loss is used for regression of the illumination-related feature vector, the smooth-L1 loss is mainly calculated by regression of a predicted gray value and a generated gray value of a label, and an ArcFace angle loss is used for regression of the face image feature loss, and the calculation formulas of the smooth-L1 loss and the ArcFace angle loss are as follows: wherein x represents a difference between a label value and a predicted value, N represents a number of batch samples, θ is an angular interval between a sample and a classification boundary, y i is a class label, j is a class boundary marker, i is a current face class, m is an angular interval, and S is a radius.

6. A feature self-decomposing face recognition system with illumination robustness, characterized in that, The system comprises: An image preprocessing unit, configured to, for a given face data set, first perform face detection, perform face alignment on the detected face, and then scale the aligned face picture to a specified size; An image label making unit, configured to calculate a gray value of the preprocessed face picture, perform gray value normalization, generate a corresponding illumination label according to the normalized gray value, and thus obtain an image data set with an illumination label; A feature self-decomposition face recognition network model constructing unit, configured to construct a feature self-decomposition face recognition network model, the feature self-decomposition face recognition network model comprising a backbone feature extraction network and a feature decomposition network, the backbone feature extraction network being configured to extract a face image feature, and the feature decomposition network being configured to decompose the extracted face image feature into an illumination-related feature vector and an illumination-independent feature vector; A network model training unit, configured to train the network model by using the image data set with the illumination label, train the backbone feature extraction network in a first stage, freeze the parameters of the trained backbone feature extraction network in a second stage, and train the feature decomposition network. The recognition unit is used for recognition by using the trained network model, and the recognition process includes face verification and face recognition. The face verification specifically includes: given two face pictures, the two face pictures are sent into the feature decomposition face recognition network model through image preprocessing, two illumination-independent feature vectors are obtained, the similarity between the two illumination-independent feature vectors is calculated, when the similarity is greater than the set threshold, it is determined that the two face pictures have the same identity information, otherwise, the two face pictures come from different people. The face recognition specifically includes: a large number of face feature vectors are stored in a feature library, the face feature vectors are illumination-independent feature vectors extracted by using the trained feature decomposition face recognition network model, when a face picture is given, the face picture is input into the feature decomposition face recognition network model to obtain a feature vector, the face picture with the maximum similarity value is found in the feature library according to the feature vector, and the identity information corresponding to the face picture with the maximum similarity value is determined. If the similarity value is greater than the set threshold, the identity information of the corresponding face in the feature library is output, otherwise, it is determined as a stranger face; The feature decomposition network receives the n-dimensional feature vector V extracted by the backbone feature extraction network, generates a weight vector W through a weight generation module in the feature decomposition network, multiplies the generated weight vector W and the n-dimensional feature vector V to obtain an illumination-dependent feature vector B, obtains a vector 1-W according to the weight vector W, and multiplies the vector 1-W and the n-dimensional feature vector V to obtain an illumination-independent feature vector F after removing the illumination information. The weight generation module is composed of two fully connected layers, an activation function, a batch normalization layer and a sigmoid layer, wherein one of the two fully connected layers is used for dimension reduction of the feature vector V, the other is used for dimension increase of the dimension-reduced feature vector V, and the sigmoid layer is used for limiting the element values of the corresponding positions in the weight vector W to be between [0, 1].

7. A feature self-decomposing face recognition system with illumination robustness, characterized in that, Comprise: At least one processor; And, The memory is in communication connection with the at least one processor; wherein The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the feature decomposition face recognition method with illumination robustness according to any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer readable storage medium stores computer executable instructions for causing a computer to execute the feature decomposition face recognition method with illumination robustness according to any one of claims 1 to 5.

Images