Image complex feature extraction method based on biological visual perception model

Abstract

The present invention discloses a method for extracting complex features from images based on a biological visual perception model. The method first simulates the visual cortex's cognitive mechanism to group the complex results obtained from the Gabor transform of the image, performing singular value decomposition and maximum value operations to obtain the main complex features. The method then performs depth mapping on the main complex features within each group. Finally, the deep complex features within different groups are recursively learned to generate image complex feature vectors. The present invention's method for extracting complex features from images based on a biological visual perception model uses a complex signal composed of real and imaginary parts to carry two-dimensional information in a single signal dimension. Complex neurons can better simulate biological neurons, increasing the generalization capability and classification accuracy of individual processing units and the entire network. Leveraging the powerful feature learning capabilities of deep learning, the C-HMAX model of the present invention is more consistent with human visual processing mechanisms, resulting in better feature extraction.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing methods, and in particular relates to an image complex feature extraction method based on a biological visual perception model. Background Art

[0002] Image feature extraction is a critical step in image recognition, object detection, and motion tracking. Feature extraction performance directly impacts subsequent processing. Existing common image feature extraction methods are all based on the real domain, but complex signals are increasingly appearing in real applications. Experimental results show that deep models using all-complex representations achieve significantly better classification accuracy and performance than deep networks using real representations. A common data processing approach involves first converting the complex information into real numbers by performing modulo processing, and then applying real-based processing methods. This approach results in the loss of some useful information, impacting the final representation performance. Theoretical analysis shows that complex numbers not only possess richer representational capabilities but also contribute to a memory retrieval mechanism that is robust to noise. Furthermore, compared to real-valued image feature extraction models, complex-valued image feature extraction models offer advantages such as easier optimization, better generalization, and faster learning. In recent years, with the increasing prevalence of complex signals in real applications, the application areas of complex feature extraction methods have gradually expanded, and image feature extraction based on complex representations has garnered increasing attention from researchers.

[0003] The existing image feature extraction method based on the HMAX model is based on the processing of Gabor transform amplitude. The Gabor transform phase contains useful information such as image edges and details but is not fully utilized; processing the real and imaginary parts (or amplitude and phase) of the Gabor transform complex features separately will reduce the complementary collaboration between the two and reduce the performance of the feature extraction model. Summary of the Invention

[0004] The purpose of the present invention is to provide an image complex feature extraction method based on a biological visual perception model, which is based on a human visual processing model and fully utilizes the advantages of image Gabor transform complex information, traditional learning and deep learning feature extraction.

[0005] The technical solution adopted by the present invention is: an image complex feature extraction method based on a biological visual perception model, which first imitates the visual cortex cognitive mechanism to group the complex results obtained by Gabor transform of the image and perform singular value decomposition and maximum value operation to obtain the main complex features, then perform depth mapping on the main complex features in the group respectively, and finally perform recursive learning on the deep complex features in different groups to generate image complex feature vectors.

[0006] The present invention is also characterized in that it comprises the following steps:

[0007] Step 1: Transform the image using a Gabor filter bank with 16 scales and 4 directions to obtain the Gabor transform response of the image;

[0008] Step 2: Use the singular value decomposition method to calculate the principal components of the Gabor transform response of the image pixels in different directions at the same scale in a 5×5 neighborhood to obtain the single-scale principal complex features; then, for the corresponding pixels of the principal complex images of two adjacent scales in the group, obtain the principal complex value corresponding to the maximum amplitude to obtain the principal complex features within the group;

[0009] Step 3: Use the extreme learning machine model based on local receptive field to perform deep mapping on the main complex features within the group to generate deep complex features within the group;

[0010] Step 4: In the recursive framework, the stacked complex autoencoder (SCAE) iteratively maps the deep complex features of different groups in turn to generate the final image complex feature vector.

[0011] Step 1 is as follows:

[0012] Step 1.1, define the Gabor filter as:

[0013]

[0014] In formula (1), z = (x, y), ||·|| represents the norm operation, k v =k max / f v , σ=2π,k max =π / 2, μ and v represent the direction and scale of the Gabor filter, respectively;

[0015] Step 1.2: Use a Gabor filter bank with 16 scales (v=0, 1, ..., 15) and 4 directions (μ=0, 1, ..., 3) to transform the image and obtain the Gabor transform response G μ，v (z):

[0016]

[0017] Step 2 is as follows:

[0018] Step 2.1: Transform the single-scale Gabor transform response of a single pixel in the image into a 5×5 neighborhood G μ，v The real part Re of (z) μ，v (k) and the imaginary part Im μ，v (k) is rewritten as the response matrix G in the following form:

[0019]

[0020] Step 2.2, the covariance matrix C of the response matrix G is:

[0021]

[0022] Step 2.3: Perform singular value decomposition on the response matrix G based on the covariance matrix C to obtain the main complex features of the single-scale Gabor transform of the pixel point:

[0023] G=USV T =UDiag[s1 s2][v1 v2] T (5)

[0024] In formula (5), U is a 100×2 matrix, V is a 2×2 matrix, S is a 2×2 diagonal matrix, the diagonal elements are singular values, and the first column of the matrix V, i.e. [v 1，1 v 1，2 ] describes the main information of Gabor transform. The main complex features and main complex feature amplitudes of pixel Gabor transform are:

[0025] GR μ，v ′(z)=v 1，1 +i·v 1，2 (6)

[0026]

[0027] Step 2.4: Compare the amplitudes of the main complex features of two adjacent scales within the group to obtain the main complex features within the group

[0028]

[0029] Step 3 is as follows:

[0030] Step 3.1: Randomly generate the initial complex weight kernel W:

[0031]

[0032] In formula (9), it is assumed that the size of the local receptive field is r×r, and K is the number of mapped feature maps;

[0033] Step 3.2: Calculate the main complex graph within the group Convolution features with random weight kernel W, assuming the input image size is d×d, then the size of the feature map is (d-r+1)×(d-r+1), and the convolution node (i, j) value of the kth feature map is M i，j，k for:

[0034]

[0035] Step 3.3, convolution feature M i，j，k Perform pooling operation to generate deep complex features within the group, and the combined node (p, q) value H in the kth pooling graph p，q，k for:

[0036]

[0037] In formula (11), e represents the distance from the pooling center to the pooling edge.

[0038] Step 4 is as follows:

[0039] Step 4.1, set group l, l = 1, ... L, L = 8 as the number of groups, H l is the image feature formed by concatenating all feature vectors of the lth group, then the first group of deep features H1 of the image is expressed by the mapping of SCAE as follows:

[0040] F1=SCAE(H1) (12)

[0041] In formula (12), F1 is the output feature value of the first iterative mapping, and SCAE(·) is the stack autoencoder operator;

[0042] Step 4.2, l-1th iteration output feature F l-1 and the feature H of scale l l Splice into a vector [F l-1 ;H l ]The input feature map to the next SCAE iteration is:

[0043] F l =SCAE([F l-1 ;H l ]), l=2,…,L (13)

[0044] Step 4.3: Concatenate the feature vectors output by all iterative mappings to form the final image complex feature vector.

[0045] The beneficial effects of the present invention are as follows: in the image complex feature extraction method based on the biological visual perception model of the present invention, the complex signal composed of the real part and the imaginary part carries two-dimensional information in a single signal dimension, and the complex neurons can better simulate biological neurons, thereby increasing the generalization ability and classification accuracy of a single processing unit and the entire network; with the help of the powerful feature learning ability of deep learning, the C-HMAX (Complex-Valued HMAX, C-HMAX) model of the present invention is more in line with the human visual processing mechanism and has a better feature extraction effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1It is a flow chart of the method for extracting complex features of an image based on a biological visual perception model of the present invention. DETAILED DESCRIPTION

[0047] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0048] Example 1

[0049] The invention provides an image complex feature extraction method based on a biological visual perception model. The similarity between the present invention and the existing HMAX method is that both follow the important role of directional information in visual perception and understanding. The main difference is that the present invention C-HMAX uses the Gabor complex feature representation mechanism and a deep complex network to perform deep feature mapping on complex feature vectors, improving the existing HMAX real feature extraction model into a complex feature extraction model, giving it the advantages of a complex feature extraction model; C-HMAX combines the Gabor filter's simulation of the 2D receptive field of simple cells in the mammalian visual cortex, traditional learning and deep learning's feature mapping capabilities, and enhances the image's expressive ability.

[0050] Example 2

[0051] like Figure 1 As shown in Figure 2, the C-HMAX method first applies a Gabor transform to the image, mimicking the cognitive mechanisms of the visual cortex. Next, the resulting complex results are subjected to group-by-group subscale SVD and group-by-group maximum operations to obtain the primary complex features within each group. Next, depth mapping is performed on the primary complex features within each group. Finally, recursive learning is performed on the deep complex features within all groups to generate the complex feature vector for the image. The specific steps for C-HMAX feature extraction are described below.

[0052] (1) Image Gabor transform

[0053] The Gabor transform of a single-channel image is a convolution operation between the image and a set of Gabor filters of different scales and directions. The definition of Gabor filter is:

[0054]

[0055] Where z = (x, y), ||·|| represents the norm operation, k v =k max / f v , σ=2π,k max =π / 2, μ and v represent the direction and scale of the Gabor filter, respectively. Here, a Gabor filter bank with 16 scales (v = 0, 1, ..., 15) and 4 directions (μ = 0, 1, ..., 3) is selected to process the image. The Gabor transform of a single-channel image I(z) is represented by the following convolution operation:

[0056]

[0057] (2) Principal component analysis of complex features

[0058] The image is transformed using a Gabor filter bank with 4 directions and 16 scales to obtain 4×16=64 response maps. These 64 response maps are divided into 8 groups, each of which includes 2 adjacent scales and all 4 direction responses in each scale.

[0059] The principal component analysis of complex features is performed on a group-by-group and scale basis. Specifically, singular value decomposition (SVD) is used to estimate the principal complex features of each pixel in a 5×5 neighborhood of the Gabor transform responses of the image pixel in different directions at the same scale. This yields a downsampled single-scale principal complex feature for all directional response maps at each scale. The principal complex value corresponding to the maximum amplitude of the corresponding pixels in the principal complex maps of two adjacent scales within the group is the final, invariant principal complex feature within the group.

[0060] First, the real part Re of the Gabor transform response of a single pixel in the image in all directions at a single scale in a 5×5 neighborhood is calculated. μ，v (k) and the imaginary part Im μ，v (k) is written as the response matrix G of the following form

[0061]

[0062] Second, the covariance matrix C of the response matrix G is

[0063]

[0064] Then, the response matrix G is subjected to singular value decomposition according to the covariance matrix C to obtain the main complex features of the single-scale Gabor transform of the pixel points.

[0065] G=USV T =UDiag[s1 s2][v1 v2] T

[0066] Where U is a 100×2 matrix and V is a 2×2 matrix. S is a 2×2 diagonal matrix with singular values ​​on the diagonal elements. The matrix also expresses the relative energy or weight value of the Gabor transform of the corresponding pixel in the real and imaginary parts. The first column of the matrix V, i.e. [v 1，1 v 1，2 ] describes the main information of Gabor transform. The main complex features GR and main complex feature amplitude GRA of the single-scale Gabor transform of the pixel point are:

[0067] GR μ，v ′(z)=v 1，1 +i·v 1，2

[0068]

[0069] Finally, the main complex feature with the largest amplitude of two adjacent scales within the group is the main complex feature of the group.

[0070]

[0071] (3) Depth mapping of main complex features within the group

[0072] The extreme learning machine model based on local receptive field is used to deeply map complex features. All parameters in the model are complex values. First, the initial complex weight kernel W is randomly generated.

[0073]

[0074] Here we assume that the local receptive field size is r×r, and K is the number of mapped feature maps. Then, we calculate the convolution features of the principal component complex map and the random weight kernel. If the input image size is d×d, then the size of the feature map is (d-r+1)×(d-r+1), and the convolution node (i, j) value of the kth feature map is M i，j，k for:

[0075]

[0076] Finally, the depth complex features are pooled to generate depth map features. Here, e represents the distance from the pooling center to the pooling edge, and the combined node (p, q) value H in the kth pooling graph is p，q，k for:

[0077]

[0078] The pooled feature map generated here has the same size as the random kernel convolution feature map, which is (d-r+1)×(d-r+1).

[0079] (4) Recursive learning of complex features between groups

[0080] The deep complex features within each group are learned using a recursive complex stacked autoencoder to form the final deep complex features of the image. The recursive stacked complex autoencoder model is composed of multiple stacked complex autoencoders (SCAEs). First, deep complex features are extracted for each group of the image. These features are then iteratively mapped to the image features using the stacked complex autoencoders (SCAEs) within a recursive framework.

[0081] For group l, l=1, ...L, L=8 is the number of groups, H l It is the image feature formed by concatenating all the feature vectors of the first group. First, the first group of deep features H1 of the image is mapped by SCAE as follows:

[0082] F1=SCAE(H1)

[0083] Here, F1 is the output feature value of the first iteration mapping, and SCAE(·) is the stacked autoencoder operator. Then, the l-1th iteration output feature F l-1 and the feature H of scale l l Splice into a vector [F l-1 ;H l ] Input to the next SCAE iteration feature map:

[0084] F l =SCAE([F l-1 ;H l ]), l=2,…,L

[0085] Finally, the feature vectors output by all iterative mappings are concatenated to form the final image representation vector.

[0086] Example 3

[0087] Fingerprint image liveness detection experiments were conducted to verify the effectiveness of the C-HMAX method. The average classification error (ACE) of C-HMAX and other methods on the LivDet2011DB, LivDet2013DB, ​​and LivDet2015DB databases is shown in the table below.

[0088] Table 1 Experimental results of various comparison methods on LivDet2011DB

[0089]

[0090]

[0091] Table 2 Experimental results of various comparison methods of LivDet2013DB

[0092]

[0093] Table 3 Experimental results of various comparison methods of LivDet2015DB

[0094]

Claims

1. An image complex feature extraction method based on a biological visual perception model, characterized in that: First, the complex results obtained by Gabor transform of the image are grouped to simulate the cognitive mechanism of the visual cortex and perform singular value decomposition and maximum value operation to obtain the main complex features. Then, the main complex features in the group are depth mapped respectively. Finally, the deep complex features in different groups are recursively learned to generate the complex feature vector of the image. The following steps are included: Step 1: Transform the image using a Gabor filter bank with 16 scales and 4 directions to obtain the Gabor transform response of the image; Step 2: Use the singular value decomposition method to calculate the principal components of the Gabor transform response of the image pixels in different directions at the same scale in a 5×5 neighborhood to obtain the single-scale principal complex features; then, for the corresponding pixels of the principal complex images of two adjacent scales in the group, obtain the principal complex value corresponding to the maximum amplitude to obtain the principal complex features within the group; Step 3: Use the extreme learning machine model based on local receptive field to perform deep mapping on the main complex features within the group to generate deep complex features within the group; Step 4: In the recursive framework, the stacked complex autoencoder (SCAE) iteratively maps the deep complex features of different groups in turn to generate the final image complex feature vector.

2. The method for extracting complex features of an image based on a biological visual perception model according to claim 1, wherein: The step 1 is specifically as follows: Step 1.1, define the Gabor filter as: In formula (1), z = (x, y), ||·|| represents the norm operation, k v =k max / f v , σ=2π,k max =π / 2, μ and v represent the direction and scale of the Gabor filter, respectively; Step 1.2: Use a Gabor filter bank with 16 scales (v = 0, 1, ..., 15) and 4 directions (μ = 0, 1, ..., 3) to transform the image and obtain the Gabor transform response G μ,v (z):

3. The method for extracting complex features of an image based on a biological visual perception model according to claim 2, wherein: The step 2 is specifically as follows: Step 2.1: Transform the single-scale Gabor transform response of a single pixel in the image into a 5×5 neighborhood G μ,v The real part Re of (z) μ,v (k) and the imaginary part Im μ,v (k) is rewritten as the response matrix G in the following form: Step 2.2, the covariance matrix C of the response matrix G is: Step 2.3: Perform singular value decomposition on the response matrix G based on the covariance matrix C to obtain the main complex features of the single-scale Gabor transform of the pixel point: <h2 style=";text-align:left;direction:ltr">G=USV<h2 style=";text-align:left;direction:ltr"> T <h2 style=";text-align:left;direction:ltr"> =UDiag[s1 s2][v1 v2]<h2 style=";text-align:left;direction:ltr"> T <h2 style=";text-align:left;direction:ltr"> (5) In formula (5), U is a 100×2 matrix, V is a 2×2 matrix, S is a 2×2 diagonal matrix, the diagonal elements are singular values, and the first column of the matrix V, i.e. [v 1,1 v 1,2 ] describes the main information of Gabor transform. The main complex features and main complex feature amplitudes of pixel Gabor transform are: GR μ,v ′(z)=v 1,1 +i·v 1,2 (6) Step 2.4: Compare the amplitudes of the main complex features of two adjacent scales within the group to obtain the main complex features within the group 4. The method for extracting complex features of an image based on a biological visual perception model according to claim 3, wherein: The step 3 is specifically as follows: Step 3.1: Randomly generate the initial complex weight kernel W: In formula (9), it is assumed that the size of the local receptive field is r×r, and K is the number of mapped feature maps; Step 3.2: Calculate the main complex graph within the group Convolution features with random weight kernel W, assuming the input image size is d×d, then the size of the feature map is (d-r+1)×(d-r+1), and the convolution node (i, j) value of the kth feature map is M i,j,k for: Step 3.3, convolution feature M i,j,k Perform pooling operation to generate deep complex features within the group, and the combined node (p,q) value H in the kth pooling graph p,q,k for: In formula (11), e represents the distance from the pooling center to the pooling edge.

5. The method for extracting complex features of an image based on a biological visual perception model according to claim 4, wherein: The step 4 is specifically as follows: Step 4.1, set group l, l = 1, ... L, L = 8 is the number of groups, H l is the image feature formed by concatenating all feature vectors of the lth group, then the first group of deep features H1 of the image is expressed by the mapping of SCAE as follows: F1=SCAE(H1) (12) In formula (12), F1 is the output feature value of the first iterative mapping, and SCAE(·) is the stack autoencoder operator; Step 4.2, l-1th iteration output feature F l-1 and the feature H of scale l l Splice into a vector [F l-1 ;H l ]The input feature map to the next SCAE iteration is: F l =SCAE([F l-1 ;H l ]),l=2,…,L (13) Step 4.3: Concatenate the feature vectors output by all iterative mappings to form the final image complex feature vector.

Images