A Robust Image Hash Authentication Method Based on Texture and Statistical Features

Abstract

The present invention discloses an image hashing authentication method based on texture and statistical features. The image hashing authentication method includes: Step S1: Preprocessing. First, convert the RGB image to the YCbCr color space and perform NSCT transformation on the Y component. Step S2: Feature extraction and representation. Extract texture features and statistical features from the NSCT sub-band images to obtain more useful image information. Among them, the texture features are obtained by mapping the NSCT low-frequency sub-band images using the uniform local binary pattern descriptor (ULBP) into histograms, and the dimension of the vector of this feature is smaller than that of the original local binary pattern descriptor LBP histogram. On the other hand, perform singular value decomposition (SVD) on the NSCT high-frequency sub-band images, and use the singular vector distance to obtain statistical features. Step S3: Combine the above two kinds of texture features and statistical features to generate an image hash and perform image authentication. The present invention has the advantages of simple principle, wide application range, good robustness and discrimination in detection performance, etc.

Description

Technical Field

[0001] The present invention mainly relates to the technical fields of image processing and information security, and in particular, to a robust image hashing authentication method based on texture and statistical features in the non-subsampled contourlet transform domain (NSCT). Background Art

[0002] With the continuous development of multimedia and Internet communication technologies, digital contents such as images, videos, and audios have gradually become carriers of information dissemination. Since digital images contain a large amount of data, are insensitive to distortion, and are easy to edit, the phenomenon of malicious tampering and destruction of digital images is relatively serious, which easily leads to copyright authentication problems. Image hashing is an effective technology in the field of digital media. It extracts a compact representation called a hash from the input image according to the visual image content and has been widely used in scenarios such as image content authentication, image tampering detection, image retrieval, image quality evaluation, image copy detection, and digital watermarking.

[0003] Existing research mainly conducts research on robust image hashing methods from four aspects. One is the robust image hashing method based on invariant feature transformation. This type of method extracts robust features from the transform domain and then generates an image hash using the transform coefficients. The advantage of these methods is that they are highly robust to image distortion and geometric attacks. Currently, transforms commonly used for image hash generation include discrete Fourier transform (DFT), discrete wavelet transform (DWT), and discrete cosine transform (DCT), etc. The second is the robust image hashing method based on statistical features. Statistical features (such as mean value, high-order moments, image intensity, and histogram, etc.) show invariance to minor changes in the image. The third is the robust image hashing method based on texture statistical features, mainly including gray-level co-occurrence matrix, etc. The fourth is the robust image hashing method based on data dimensionality reduction. In this type of method, data dimensionality reduction maps high-dimensional data to a low-dimensional space and is an effective data compression technology. Image hashing is a compact representation of an image, and the process of hash extraction is essentially a process of data compression. Commonly used data dimensionality reduction algorithms include singular value decomposition (SVD), local linear embedding (LLE), non-negative matrix factorization (NMF), principal component analysis (PCA), and multi-dimensional scaling (MDS), etc.

[0004] Generally, a good image hash satisfies two basic performances. One performance is robustness, that is, two images with similar visual contents should have the same or similar hash values. The other basic performance is discriminability, which means that images with different visual contents should have completely different hash values. Currently, most image hash technologies cannot achieve good classification performance between robustness and discriminability. Non-Subsampled Contourlet Transform (NSCT) is one of the multi-scale transform methods frequently used in recent years and has been widely used in image processing, such as image segmentation, image fusion, digital watermarking, image enhancement, etc. The non-subsampled contourlet transform (NSCT) is a two-dimensional image transform technology that retains the multi-scale decomposition, multi-directionality, and anisotropy of the contourlet transform, and cancels the downsampling and upsampling operations in the process of decomposing and reconstructing image information, making the size of each sub-band image exactly the same as that of the original image. NSCT consists of two parts: a non-subsampled pyramid filter bank (Non-subsampled Pyramid, NSP) and a non-subsampled directional filter bank (Non-subsamples Directional Filter Banks, NSDFB). Among them, NSP realizes the multi-resolution decomposition of the image, and NSDFB realizes the multi-directional decomposition of the image.

[0005] However, how to use the NSCT image hash algorithm has not been well studied. To address this problem, the present invention proposes an image hash method based on the combination of texture features and vector distance statistical features. This method decomposes the low-frequency sub-band image and the high-frequency sub-band image from the YCbCr color space through NSCT. The uniform binary pattern is used to extract texture features from the low-frequency image, and the high-frequency image is decomposed by singular value decomposition, and the image features are compressed by calculating the distance of singular vectors. Summary of the Invention

[0006] The technical problem to be solved by the present invention is: aiming at the problems existing in the prior art, the present invention provides a robust image hash authentication method with a simple principle, a wide application range, and a good balance between robustness and discriminability.

[0007] To solve the above problems, the present invention adopts the following technical solutions:

[0008] A robust image hash authentication method based on texture and statistical features, which includes:

[0009] Step S1: Preprocessing. First, convert the RGB color image to the YCbCr color space, and perform the NSCT non-subsampled contour transform on the Y color component.

[0010] Step S2: Feature extraction and representation. Texture features and statistical features are extracted from the NSCT sub-band images to obtain more useful information of the images. Among them, the texture features are obtained by mapping the uniform local binary pattern descriptor (ULBP) from the NSCT low-frequency sub-band image into a histogram. This feature has a smaller vector dimension than the histogram of the original local binary pattern descriptor LBP. On the other hand, singular value decomposition (SVD) is performed on the NSCT high-frequency sub-band image, and statistical features are obtained using the singular vector distance.

[0011] Step S3: Combine the above two kinds of texture features and statistical features to generate an image hash and perform image authentication.

[0012] As a further improvement of the method of the present invention: in the step S1, in order to reduce the influence of the content retention operation on the input image, first, the input image is converted to a unified size M×M using bilinear interpolation operation. This operation can make the constructed hash resistant to image scaling and generate a hash sequence of unified length. Then, a Gaussian low-pass filter with a standard deviation of 1 for a 3×3 convolution mask is performed on the resized image. This operation can reduce the influence of noise on the image. Finally, the filtered color image is converted to the YCbCr color space. The YCbCr color space has consistency with human visual perception, and the luminance component Y is used to represent the input image.

[0013] As a further improvement of the method of the present invention: in the step S1, first, the image is decomposed by a non-subsampled pyramid filter bank NSP to obtain a low-frequency sub-band and a high-frequency sub-band. Then, a non-subsampled directional filter bank NSDFB transform is performed on the high-frequency sub-band image output by the NSP to achieve multi-directional decomposition. After levels of NSCT decomposition, finally, sub-band images consistent with the size of the input image can be obtained, where is the decomposition level under the decomposition scale, and the decomposition scale is [0, 1, 2].

[0014] As a further improvement of the method of the present invention: in the step S2, in order to effectively apply the ULBP descriptor to extract local texture features, the following steps are performed:

[0015] First, extract the NSCT low-frequency sub-band image , and divide the low-frequency sub-band image of size M×M into a series of image sub-blocks of size m×m. Therefore, the number of blocks is , and the blocking method is as follows:

[0016] ;

[0017] Then, calculate the ULBP of each image block to obtain a ULBP map of size ​​ , the mapping consists of decimal values. It should be noted that the present invention does not consider the four boundary pixels of an image block as the central pixel, so , then, use to represent 's histogram, and the derivation formula is as follows:

[0018] ;

[0019] ;

[0020] ULBP has fewer dimensions, more texture details, and better robustness compared to traditional LBP. After obtaining the ULBP mapping 's histogram , calculate 's histogram statistical moments, and the calculation formula is as follows:

[0021] ;

[0022] ;

[0023] Among them, N is the number of variables on the horizontal axis of the histogram, n is the variable on the horizontal axis of the histogram, and the value range is [1, 2,.., 59], is the probability distribution of n, is the histogram statistical value at n. Connect the histogram statistical moments of the image block to construct the local texture feature .

[0024] As a further improvement of the method of the present invention: in the step S2, in order to obtain more image details, the present invention learns stable image structure information from the high-frequency images obtained by NSCT decomposition; first, perform singular value decomposition (SVD) on the NSCT high-frequency subband images; use the singular vector distance to obtain statistical features; first, for each NSCT high-frequency subband image , perform the following SVD operation;

[0025] ;

[0026] Among them, is a diagonal matrix of size M×M, with M singular values, and are two orthogonal matrices corresponding to of size M×M. Next, for each high-frequency subgraph, the present invention collects and 's first largest singular vectors (and corresponding to the largest singular vectors of singular values), as the feature matrix. Among them, and are respectively and 's -th vector, with size M×1. Therefore, the feature matrix , with size M×D, D = 。

[0027] As a further improvement of the method of the present invention: in the step S2, in order to obtain compact hashing, each column of the feature matrix is regarded as a vector, and these feature vectors are used to construct the hash; the present invention selects the vector distance as the feature, and the calculation details of the vector distance are as follows: First, for the feature matrix of each high-frequency subgraph, the reference vector is as follows:

[0028] ;

[0029] ; ;

[0030] Among them, is the 's -th element, is the 's -th element; Second, use the norm to calculate the distance between the reference vector and the vector ,

[0031] ;

[0032] Finally, connect the vector distances to construct the distance feature. Since each high-frequency subgraph has D vector distances and there are L2 high-frequency subgraphs in total, the distance feature matrix is as follows:

[0033] ;

[0034] where the size of Q is × , stretch the feature matrix Q into a one-dimensional sequence .

[0035] As a further improvement of the method of the present invention: in the step S3, the present invention quantizes the local texture feature and the singular vector distance feature to generate the final image hash sequence as follows:

[0036] ;

[0037] ;

[0038] where represents the th element of the texture feature represents the th element of the statistical feature and respectively represent the and means, the represents the th element of and are concatenated to obtain the final hash , so the hash length of the method of the present invention is bits.

[0039] Compared with the prior art, the advantages of the present invention are as follows:

[0040] The overall architecture of the method of the present invention includes three modules: image preprocessing, image feature extraction and representation, and hash quantization and generation. The present invention proposes a new image hashing algorithm based on texture and statistical features in the NSCT domain. Compared with the prior art, it can achieve a good balance between robustness and discrimination ability. In the preprocessing module, in order to prove that the YCbCr color space is the best color space choice for the proposed method, three other color spaces are used to design the hashing algorithm, namely the CIE Lab color space, the HSI color space, and the HSV color space. The results show that the YCbCr color space has the best hashing performance. In the feature extraction module, the present invention adopts the NSCT transform domain. By using the characteristics of NSCT, the sub-band images in different scales and different directions are decomposed. Extracting features from the obtained sub-band images can improve the robustness of the method of the present invention. At the same time, the ULBP descriptor is executed on the NSCT low-frequency sub-band image and mapped to a histogram. The image is represented by the histogram statistical moment as the texture feature. Compared with the original LBP histogram, the ULBP histogram has fewer vector dimensions and can represent the image more compactly. When performing statistical feature extraction and representation, the present invention uses singular value decomposition (SVD) on the NSCT high-frequency sub-band image, calculates the singular vector distance to construct the structural feature. Since the vector distance is stable for common operations and the storage space required for the vector distance is not large, it helps to generate a compact image hash. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1It is a schematic flow chart of the method of the present invention;

[0042] Figure 2 It is a schematic diagram of the implementation principle of the present invention in a specific application example. Specific embodiments

[0043] The present invention will be further described in detail below with reference to the accompanying drawings of the specification and specific examples.

[0044] As Figure 1 and Figure 2 shown, a robust image hashing authentication method based on texture and statistical features of the present invention extracts texture features and statistical features in the low-frequency sub-band and high-frequency sub-band respectively in the NSCT transform domain, and combines these two features to obtain a short authentication image hash that achieves a good balance between robustness and discriminability. The process of the method of the present invention includes:

[0045] Step S1: Preprocessing. First, convert the RGB color image to the YCbCr color space, and perform the NSCT non-subsampled contour transform on the Y color component.

[0046] Step S2: Feature extraction and representation. Extract texture features and statistical features from the NSCT sub-band image to obtain more useful information of the image; among them, the texture features are obtained by mapping the NSCT low-frequency sub-band image using the uniform local binary pattern descriptor (ULBP) as a histogram, and the dimension of the vector of this feature is smaller than that of the original local binary pattern descriptor LBP histogram; on the other hand, perform singular value decomposition (SVD) on the NSCT high-frequency sub-band image, and use the singular vector distance to obtain statistical features;

[0047] Step S3: Combine the above two texture features and statistical features to generate an image hash and perform image authentication.

[0048] In a specific application example, the present invention further optimizes the preprocessing process, and the specific optimization is as follows:

[0049] To reduce the influence of the content retention operation on the input image, first use the bilinear interpolation operation to convert the input image to a unified size of M×M. This operation can make the constructed hash resistant to image scaling and generate a hash sequence of a unified length; then, perform Gaussian low-pass filtering with a standard deviation of 1 on the resized image using a 3×3 convolution mask. This operation can reduce the influence of noise on the image; finally, convert the filtered color image to the YCbCr color space. The YCbCr color space has consistency with human visual perception, and the luminance component Y is used to represent the input image;

[0050] In a specific application example, as a further improvement of the present invention: First, the image is decomposed by a non-downsampled pyramid filter bank (NSP) to obtain a low-frequency sub-band and a high-frequency sub-band. Then, a non-downsampled directional filter bank (NSDFB) transform is performed on the high-frequency sub-band image output by the NSP to achieve multi-directional decomposition. After levels of NSCT decomposition, finally, sub-band images with the same size as the input image can be obtained, where is the decomposition level and is the decomposition scale at level

[0051] In a specific application example, as a further improvement of the present invention: To effectively apply the ULBP descriptor to extract local texture features, the following steps are performed:

[0052] First, the NSCT low-frequency sub-band image is extracted. The low-frequency sub-band image of size M×M is segmented into a series of image sub-blocks of size m×m. Therefore, the number of blocks is , and the blocking method is as follows:

[0053] ;

[0054] Then, the ULBP of each image block is calculated to obtain a ULBP map of size , which consists of decimal values. It should be noted that in the present invention, the four boundary pixels of the image block are not regarded as the central pixel. Therefore . Then, is used to represent 's histogram, and the derivation formula is as follows:

[0055] ;

[0056] ;

[0057] The ULBP has fewer dimensions, more texture details, and better robustness compared to the traditional LBP. After obtaining the histogram of the ULBP map of the image block, the histogram statistical moments of are calculated, and the calculation formula is as follows:

[0058] ;

[0059] ;

[0060] where N is the number of variables on the horizontal axis of the histogram, n is the variable on the horizontal axis of the histogram, and the value range is [1, 2,.., 59]. The probability distribution for n is the histogram statistical value at n. Connect the histogram statistical moments of the image patches to construct local texture features .

[0061] In a specific application example, as a further improvement of the method of the present invention: in step S2, in order to obtain more image details, the present invention learns stable image structure information from the high-frequency images obtained by NSCT decomposition. First, perform singular value decomposition (SVD) on the NSCT high-frequency subband images, and obtain statistical features using the singular vector distance; first, for each NSCT high-frequency subband image , perform the following SVD operation;

[0062] ;

[0063] Among them, is a diagonal matrix of size M×M, with M singular values, and are two orthogonal matrices corresponding to of size M×M; Next, for each high-frequency subgraph, the present invention collects and of the first singular vectors (and the singular vectors corresponding to the largest singular values), as the feature matrix; Among them, and are respectively and of the th vector, of size M×1, so the feature matrix , of size M×D, D = 。

[0064] In a specific application example, as a further improvement of the method of the present invention: in step S2, in order to obtain compact hashing, each column of the feature matrix is regarded as a vector, and these feature vectors are used to construct hashing; the present invention selects vector distance as the feature because vector distance is robust to numerical operations and can ensure the compactness of the hash sequence. The calculation details of the vector distance are as follows: First, for the reference vector of the feature matrix of each high-frequency subgraph are as follows:

[0065] ;

[0066] ;

[0067] Among them, is of -th element, is of -th element; Secondly, use norm to calculate the distance between the reference vector and the vector ,

[0068] ;

[0069] Finally, connect the vector distances to construct the distance feature. Since each high-frequency subgraph has D vector distances and there are L2 high-frequency subgraphs in total, the distance feature matrix is as follows:

[0070] ;

[0071] where the size of Q is × , stretch the feature matrix Q into a one-dimensional sequence .

[0072] In a specific application example, as a further improvement of the method of the present invention: in the step S3, the present invention quantizes the local texture feature and the singular vector distance feature to generate the final image hash sequence as follows:

[0073] ;

[0074] ;

[0075] where represents the -th element of the texture feature , represents the -th element of the statistical feature , and respectively represent and means, represents the -th element of, represents the -th element of and are connected to obtain the final hash , so the hash length of the method of the present invention is bits.

[0076] The above are only the preferred implementation embodiments of the present invention. The protection scope of the present invention is not limited to the above embodiments. Any technical solutions falling within the idea of the present invention belong to the protection scope of the present invention. It should be noted that for those of ordinary skill in the art, several improvements and refinements made without departing from the principle of the present invention should be regarded as within the protection scope of the present invention.

Claims

1. An image hashing authentication method based on texture and statistical features, characterized in that the steps Including: Step S1: Preprocessing. First, convert the RGB color image to the YCbCr color space, and perform the NSCT (Non-Subsampled Contourlet Transform) on the Y color component. Step S2: Feature extraction and representation. Extract texture features and statistical features on the NSCT subband images to obtain useful information of the image. Among them, the texture features are obtained by mapping the NSCT low-frequency subband image using the Uniform Local Binary Pattern descriptor (ULBP) as a histogram. This feature has a smaller vector dimension than the histogram of the original Local Binary Pattern descriptor LBP. First, extract the NSCT low-frequency subband image I0, and divide the low-frequency subband image of size M×M into a series of image sub-blocks of size m×m. Therefore, the number of blocks is L1 = (M×M) / (m×m). The blocking method is as follows: After performing the blocking, calculate the ULBP of each image block to obtain the ULBP map \(W\) of size \(m'\times m'\). i (where \(i = 1, 2,\cdots, L1\)), the map consists of decimal values; the four boundary pixels of the image block are not regarded as the central pixel, so \(m'=m - 2\). Then, use \(U\) i (where \(i = 1, 2,\cdots, L1\)) to represent the histogram of \(W\) i The derivation formula is as follows: Among them, I C represents the central pixel of the image block, and ULBP P，radius (I c ) represents the UniformLBP mapping distributed within the P neighborhood with a radius of radius. Z(u, v) defines a mathematical expression of an identity matrix; obtain the histogram U i of the ULBP mapping W i of the image block. After that, calculate the histogram statistical moments of U i . The calculation formula is as follows: Among them, N is the number of variables on the horizontal axis of the histogram, n is the variable on the horizontal axis of the histogram, and the value range is [1, 2,.., 59], p i (n) is the probability distribution of n, h i (n) is the histogram statistical value at n. Connect the histogram statistical moments of the image blocks to construct the local texture feature F1 = (h l (1), h l (2),..., h l (L1)], where L1 represents the length of the texture feature; On the other hand, perform singular value decomposition (SVD) on the NSCT high-frequency subband images, and obtain statistical features by using the singular vector distance; first, for each NSCT high-frequency subband image I j (j = 1, 2,.., L2), perform the following SVD operation: I j = U j · S j · V j , Among them, S j is a diagonal matrix of size M×M with M singular values, U j and V j are two orthogonal matrices of size M×M corresponding to I J ; Next, for each high-frequency subgraph, collect the first maximum ξ singular vectors of U j and V j ​ [u j (1) , u j (2) ,..., u j (ξ) , v j (1) , v j (2) ,..., v j (ξ) , 1 ≤ j ≤ L2 as the feature matrix; where, u j (λ) and v j (λ) (λ = 1, 2,..., ξ) are the λ-th vectors of U j and V j respectively, with size M × 1. Thus, the feature matrix is Γ j = [r j (1) , r j (2) ,..., r j (ξ) , r j (ξ+1) , r j (ξ+2) ,..., r j (ξ+ξ) , with size M × D, D = ξ + ξ; To obtain a compact hash, each column of the feature matrix Γ j is regarded as a vector, and these feature vectors are used to construct the hash; the vector distance is selected as the feature, and the calculation details of the vector distance are as follows: First, for the feature matrix Γ j of each high-frequency subgraph, the reference vector R j is as follows: R j = [R j (1), R j (2),..., R(M)] T , where R j (a) is the a-th element of R j , and is 's a-th element; secondly, use the L2 norm to calculate the distance between the reference vector R j and the vector . Finally, construct the distance feature by connecting vector distances. Since each high-frequency subgraph has D vector distances and there are L2 high-frequency subgraphs in total, the distance feature matrix is as follows: where the size of Q is ×D×L2, and stretch the feature matrix Q into a one-dimensional sequence F2 = [Q1, Q2,..., Q D×L2 ; Step S3: Combine the above texture feature F1 and statistical feature F2 to generate an image hash and perform image authentication. For the local texture feature F1 = [h l (1), h l (2),..., h l (L1)] and the singular vector distance feature F2 = [Q1, Q2,..., Q D×L2 , perform quantization to generate the final image hash sequence as follows: where h l (L) represents the L-th element of the texture feature F1, Q(H) represents the H-th element of the statistical feature F2, mean(F1) and mean(F2) respectively represent the means of F1 and F2, H1(L) represents the L-th element of H1, H2(H) represents the H-th element of H2, and connecting H1 and H2 gives the final hash Hash = [H1, H2].

2. The image hashing authentication method based on texture and statistical features according to claim 1, wherein In the preprocessing of step S1, first use the bilinear interpolation operation to convert the input image to a unified size of M×M. This operation can make the hash resistant to image scaling and generate a hash sequence of unified length. Then, perform Gaussian low-pass filtering with a 3×3 convolution mask and a standard deviation of 1 on the resized image. This operation can reduce the influence of noise on the image. Finally, convert the filtered color image to the YCbCr color space. The YCbCr color space has consistency with human visual perception, and the luminance component Y is used to represent the input image.

3. The method for texture- and statistical feature-based image hashing authentication according to claim 1, wherein In the feature extraction and representation step S2, the image is first decomposed by a non-subsampled pyramid filter bank NSP to obtain a low-frequency sub-band and a high-frequency sub-band, and then a non-subsampled directional filter bank NSDFB transform is performed on the high-frequency sub-band image output by the NSP to achieve multi-directional decomposition; After level NSCT decomposition, we can finally get The subband image has the same size as the input image, where l k is the decomposition scale under the decomposition level k, and the decomposition scale is [0,1,2].

Images