A method for evaluating the visual security of selectively encrypted images

Abstract

The application provides a visual security evaluation method for selectively encrypting images. The method comprises: detecting a plaintext image P and a corresponding encrypted image E respectively to obtain a feature point set K1 of the plaintext image P and a feature point set K2 of the encrypted image E; calculating gradient amplitude maps of the plaintext image P and the encrypted image E respectively; performing feature matching on K1 and K2 by using the minimum Euclidean distance, and recording the matched feature point set as K3; calculating a structural similarity map between the gradient amplitude maps of the plaintext image P and the encrypted image E, and obtaining an image structural feature similarity S between the image P and the image E according to the structural similarity map; calculating an image content perception feature similarity S between the image P and the image E according to K1 and K3; constructing a support vector regression model and training to obtain an image visual security evaluation model; inputting the feature similarity vectors [S T ,S match ] of the image P and the image E into the trained image visual security evaluation model to obtain an image visual security evaluation result. T match T match

Description

TECHNICAL FIELD

[0001] The present application relates to the field of image and video processing and computer vision, and particularly relates to a visual security evaluation method for selectively encrypted images. BACKGROUND

[0002] With the rapid development of 5G technology, cloud storage and network communication technology, data has become more and more easy to obtain and use, and digital images have become an important carrier for people to transmit information on the network due to their more intuitive data expression. In order to protect the security of images, researchers have developed various image security technologies, such as digital watermarking, steganography, encryption, etc., and encryption is the most widely used method at present, which can guarantee the security and integrity of data. The existing image encryption methods can be roughly divided into two categories: full encryption and selective encryption. Full encryption means that the entire image is encrypted, and we cannot obtain any information about the original image from the encrypted image. In practical applications, when faced with real-time transmission of massive data, traditional full encryption methods are not suitable. Therefore, researchers have proposed selective encryption algorithms. Compared with full encryption algorithms, selective encryption has two main advantages. First, because only a part of the data needs to be encrypted, the encryption and decryption speed is very fast, so it performs quite well in real-time communication processing of massive image data. Second, the encryption strength can be adjusted according to needs, preventing the basic visual properties of the original data from being misused while enhancing the usability of the ciphertext image.

[0003] For selective encryption algorithms, in addition to the security against hostile attacks, the encrypted content must be difficult for human perception to understand, the latter requirement is called visual security, which refers to quantitatively evaluating the visual content of the ciphertext image to measure the performance of the selective encryption method, so as to help us optimize the encryption algorithm, while we can flexibly choose the encryption strength according to the needs, adjust the availability and privacy of the ciphertext image. At present, the visual security evaluation index (VSI) of selective encryption image can be divided into two categories. One is the image quality assessment (IQA), which is motivated by the belief that there is a strong relationship between image quality and visual security. Poor quality images often have high visual security. This category of indicators focuses on evaluating whether the encrypted image achieves semantic security, that is, whether the attacker can obtain any information from the encrypted image. Selective encryption images cannot achieve semantic security. These image quality assessments often appear inconsistent between objective results and subjective perception when evaluating the visual security of selective encryption images. The other is the VSI specially designed for selective encryption, which extracts image features from the ciphertext image and maps them to the corresponding visual security score by analyzing the human visual system (HVS) perception mechanism. Although the existing specially designed VSI usually has better performance than the IQA method, there is still no general analysis strategy that can cover multiple categories of image encryption algorithms.

[0004] The first problem is that the visual features of encrypted images are not fully considered. Most of the current indicators do not consider the role of the information selection strategy of the human visual system in visual security evaluation. These problems will lead to inaccurate image visual security of the visual security index. Xiang et al. (Xiang T, Guo S, Li X, Perceptual visual security index based on edge and texture similarities, IEEE Transactions on Information Forensics and Security, 11.5: (2016) 951-963.) proposed a visual security evaluation method by calculating the edge and texture similarity of the plaintext image and the ciphertext image. Although it has better performance compared with existing visual security evaluation methods, they do not consider the visual importance of the image, and treat all extracted features equally, which leads to inconsistent objective results and subjective perception.

[0005] The second problem is that the extracted features are not effective. Existing visual security indicators often use image contours, textures, and brightness as features to represent image content. These features are often unstable, and the encrypted features will change significantly. Wu et al. (Wu Z, Zhang K, Ren Y, et al. Visual Security Assessment via Saliency-Weighted Structure and Orientation Similarity for Selective Encrypted Images, Security and Communication Networks, 2021 (2021) 1-16.) proposed a method for assessing the visual security of selectively encrypted images based on saliency-weighted structure and orientation similarity. This method is mainly based on structural information to design indicators. These local features are not sufficient to effectively represent important visual information. The Arnold cat map and chaotic pseudo-random generator in the encryption process will cause significant changes in image features, resulting in low structural similarity, while important visual information remains identifiable, leading to a situation where the objective result is more secure than the subjective perception of security. SUMMARY

[0006] To address the above two problems existing in the existing method, which leads to a low consistency between the objective result and the subjective perception of the visual security evaluation, the present application provides a stable and efficient method for evaluating the visual security of selectively encrypted images.

[0007] The present application provides a method for evaluating the visual security of selectively encrypted images, comprising:

[0008] Step 1: Detect the key points in the plaintext image P and the corresponding encrypted image E respectively, and calculate the feature vector of each key point in the plaintext image P and the encrypted image E respectively, to obtain the feature point set K1 of the plaintext image P and the feature point set K2 of the encrypted image E;

[0009] Step 2: Calculate the gradient amplitude map of the plaintext image P and the gradient amplitude map of the encrypted image E respectively;

[0010] Step 3: Perform feature matching on the feature point set K1 and the feature point set K2 using the minimum Euclidean distance, and record the matched feature point set as K3;

[0011] Step 4: Calculate the structural similarity map between the gradient amplitude map of the plaintext image P and the gradient amplitude map of the encrypted image E, and calculate the image structural feature similarity S between the plaintext image P and the encrypted image E based on the structural similarity map T ;

[0012] Step 5: calculating the image content perceptual feature similarity S between the plaintext image P and the encrypted image E according to the feature point sets K1 and K3 match ;

[0013] Step 6: constructing a support vector regression model and training to obtain an image visual security evaluation model;

[0014] Step 7: inputting the feature similarity vectors [S T ,S match ] of the plaintext image P and the encrypted image E into the trained image visual security evaluation model to obtain an image visual security evaluation result.

[0015] Further, in step 1, the SIFT algorithm is used as a feature point detection method to detect key points in the image and calculate the feature vectors of the key points.

[0016] Further, step 2 specifically includes:

[0017] The gradient magnitude map GM P of the plaintext image P and the gradient magnitude map GM E of the encrypted image E are calculated according to formulas (8) and (9) respectively.

[0018]

[0019]

[0020] where "*" represents a linear convolution operator, M k , k = 1, 2, 3, 4 represent four direction gradient operators, P(i, j) represents the pixel at the i-th row and j-th column position in the plaintext image P, E(i, j) represents the pixel at the i-th row and j-th column position in the encrypted image E, and mean2 represents an average value function.

[0021] Further, step 4 specifically includes:

[0022] First, the structural similarity map S GM (i, j) between the gradient magnitude map of the plaintext image P and the gradient magnitude map of the encrypted image E is calculated according to formula (10); then the average value of the structural similarity map S GM (i, j) is obtained to obtain the image structure feature similarity S T ;

[0023]

[0024] where R is a normal number to avoid instability when the denominator converges to zero.

[0025] Further, step 5 specifically includes:

[0026] Step 5.1: Calculate the foreground segmentation threshold Th of the plaintext image P according to formula (11) by using the maximum inter-class variance method;

[0027] Th = p1(m1-mG) 2 + p2(m2-mG) 2 (11)

[0028] Wherein, m1 is the average gray value of the foreground, p1 is the ratio of the foreground pixel points to all pixel points of the image, m2 is the average gray value of the background, p2 is the ratio of the background pixel points to all pixel points of the image, and mG is the global gray mean value of the image;

[0029] Step 5.2: Filter the background matching pairs in the feature point set K3 by using the foreground segmentation threshold Th, and record the number of remaining matching pairs as NOM; filter the background matching pairs in the feature point set K1 by using the foreground segmentation threshold Th, and record the number of remaining matching pairs as NOP, and calculate the image content-aware feature similarity S according to formula (12) match ;

[0030]

[0031] Further, step 6 specifically comprises:

[0032] Step 6.1: Prepare the training data set, specifically including: collect n plaintext images, and encrypt the n plaintext images to obtain corresponding n encrypted images by using different types of selective encryption algorithms; wherein, one plaintext image and its corresponding encrypted image constitute an image pair, and n image pairs constitute the training data set;

[0033] Step 6.2: Extract the image content-aware feature similarity S match and the image structure feature similarity S T of each image pair according to steps 4 and 5 respectively, and form a feature similarity vector [S T , S match ], thereby obtaining the feature similarity vector set H = {H1, H2, …, H n} of the training data set, H n represents the feature similarity vector of the nth image pair;

[0034] Step 6.3: Collect the mean opinion score MOS of all image pairs, and construct to obtain the set S = {S1, S2, …, S n}, S n represents the mean opinion score MOS of the nth image pair;

[0035] Step 6.4: According to the set H and the set S, the implicit relationship between each feature similarity vector and the corresponding mean opinion score MOS is established by using formula (13), and an image visual security evaluation model Pr_Model is obtained.

[0036] Pr_Model = SVR train(H,S) (13)

[0037] Wherein, SVR represents a support vector regression model, and a radial basis function kernel is used as a kernel function.

[0038] The beneficial effects of the present application are as follows:

[0039] (1) The present application proposes image content perception feature similarity and image structure feature similarity as selective encryption image visual security evaluation indexes based on image content perception features and image multi-directional structure features, the image content perception feature is a stable and human visual perception consistent local feature in the selective encryption image, and can represent the leakage degree of image content; the image multi-directional structure feature is a spatial structure information highly consistent with human visual perception, and can represent the change degree of image overall skeleton; therefore, based on the above two indexes as evaluation indexes, the objective result and subjective perception consistency of visual security evaluation can be ensured; finally, the support vector regression model is used to establish the relationship between the above two evaluation indexes and the subjective evaluation score, so that the prediction model of image visual security evaluation is obtained, and the accuracy of result prediction is improved.

[0040] (2) The multi-directional gradient operator is used to obtain the structure graph of the plaintext image and the encrypted image, the distortion of the image edge region is captured more accurately, and the mean value of the structure similarity graph of the image is calculated as the global structure feature similarity between images; meanwhile, considering the visual characteristics of the human eye, the maximum inter-class variance method is used to remove the background matching feature points when calculating the content perception feature similarity, and finally the ratio of the number of key point matches to the number of foreground feature points of the original image is obtained as the content perception feature similarity of the image. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 A flowchart of a selective encryption image visual security evaluation method provided by the embodiment of the present application is shown in the figure;

[0042] Figure 2 A detection diagram of local extreme points provided by the embodiment of the present application is shown in the figure;

[0043] Figure 3 A gradient operator diagram of four directions provided by the embodiment of the present application is shown in the figure;

[0044] Figure 4The confidence evaluation result of all VSIs on the IVC-SelectEncrypt database is provided for the embodiment of the present application. DETAILED DESCRIPTION

[0045] To make the objects, technical solutions and advantages of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0046] Embodiment 1

[0047] As shown in the figure, the present embodiment provides a visual security evaluation method for selectively encrypting images, comprising the following steps: Figure 1

[0048] S101: Detecting the plaintext image P and the corresponding encrypted image E respectively to extract the key points in the images, and calculating the feature vectors of each key point in the plaintext image P and the encrypted image E respectively, so as to obtain the feature point set K1 of the plaintext image P and the feature point set K2 of the encrypted image E;

[0049] Specifically, the SIFT algorithm is used as the feature point detection method to detect the key points in the images and calculate the feature vectors of the key points in the present embodiment.

[0050] Taking the feature point detection process of the plaintext image P as an example, the specific steps are as follows:

[0051] S1011: According to the scale space theory, using the Gaussian function G(i,j,σ) to process the input image P(i,j) according to formula (1) to generate the Gaussian smoothing image L(i,j,σ) and construct the Gaussian pyramid.

[0052] L(i,j,σ)=G(i,j,σ)*P(i,j) (1)

[0053]

[0054] Where (i,j) is the spatial coordinates, and "*" is the convolution operator. G(i,j,σ) is the Gaussian function, and σ is called the factor of the scale space, i.e. the blur coefficient.

[0055] S1012: After obtaining the Gaussian pyramid, the Gaussian difference of gold pyramid (DOG) is constructed using the Gaussian pyramid to detect the feature points of the image; wherein the Gaussian difference of gold pyramid is obtained by the difference of L(i,j,σ) with different σ, which is defined in mathematics as: ​

[0056]

[0057] where k is a constant.

[0058] S1013: In the difference-of-Gaussian pyramid, the local extreme value in the scale space is calculated by comparing each pixel point (marked as X) in the image with its 8 pixel values in the same scale and 9 pixel values in the adjacent upper and lower two layers, if the value is the maximum or minimum value compared with other 26 values, then the point is marked as a local extreme point, as shown in the following formula: Figure 2 Figure 2

[0059] It should be noted that the local extreme point in step S1013 is a local extreme point in discrete space rather than a local extreme point in continuous space, so further accurate positioning is needed.

[0060] S1014: In order to accurately position the feature point, using the local extreme point X0(x0,y0,σ0) in the discrete space, the Taylor ternary second-order expansion is made at X0 according to the following formula to obtain f(X), and then the derivative of f(X) is taken and set to 0 to obtain

[0061]

[0062] where represents the relative displacement amount.

[0063]

[0064] When the relative displacement amount is greater than or equal to a preset threshold, the local extreme point X0(x0,y0,σ0) is a true local extreme point, otherwise the point is regarded as a low-contrast point and is discarded.

[0065] S1015: Assign a direction to the key point according to the properties of the image; wherein the key point is the local extreme point in step S1014.

[0066] Take the key point as the center, and count the gradient amplitude M and the gradient direction θ of the pixels in the circle with a radius of 1.5 times of the scale σ of the Gaussian image where the key point is located, and draw the gradient direction histogram of the key point.

[0067]

[0068]

[0069] ​​The highest peak value in the histogram and other peak values within 80% of the highest peak value height are identified as the main direction of the key point, respectively. According to the main direction of the key point, 16 sub-regions are divided around, and the sub-regions are usually 4x4. Then the gradient amplitude in 8 directions in the sub-region is counted to obtain a 128-dimensional feature vector of each key point, thereby obtaining a feature point set K1 corresponding to the plaintext image P;

[0070] Similarly, the feature point set K2 corresponding to the ciphertext image E can be obtained according to the above process.

[0071] S102: Calculate the gradient amplitude graph of the plaintext image P and the gradient amplitude graph of the encrypted image E, respectively;

[0072] S103: Perform feature matching on the feature point set K1 and the feature point set K2 using the minimum Euclidean distance, and record the matched feature point set as K3;

[0073] Specifically, in the two feature point sets K1 and K2, feature matching is performed using the minimum Euclidean distance, the minimum distance found is multiplied by 1.5, and then compared with other distances again. If the multiplied distance is still smaller than all other distances, it is considered that the feature point pair is matched.

[0074] S104: Calculate the structural similarity map between the gradient amplitude graph of the plaintext image P and the gradient amplitude graph of the encrypted image E, and calculate the image structure feature similarity S between the plaintext image P and the encrypted image E according to the structural similarity map T ;

[0075] S105: Calculate the image content perceptual feature similarity S between the plaintext image P and the encrypted image E according to the feature point set K1 and K3 match ;

[0076] S106: Construct a support vector regression model and train to obtain an image visual security evaluation model;

[0077] S107: Input the feature similarity vectors [S T ,S match ] of the plaintext image P and the encrypted image E into the trained image visual security evaluation model to obtain the image visual security evaluation result.

[0078] This invention provides a visual security evaluation method for selectively encrypted images. Based on image content-aware features and multi-directional structural features, it proposes image content-aware feature similarity and image structural feature similarity. Image content-aware features are stable local features in a selected encrypted image that conform to human visual perception, representing the degree of leakage of image content. Multi-directional structural features are spatial structural information highly consistent with human visual perception, representing the degree of change in the overall image skeleton. A support vector regression model is used to establish the relationship between these two evaluation indicators and subjective evaluation scores, thereby obtaining a predictive model for image visual security evaluation and improving the accuracy of result prediction.

[0079] Example 2

[0080] Based on the above embodiment 1, in order to further improve the accuracy of prediction, the visual security evaluation method for selectively encrypted images provided in this embodiment starts from the perspective of extracting more accurate image content perception feature similarity and image structure feature similarity from the image. The following feature extraction method is designed, and the rest of the process is the same as in the above embodiment 1, and will not be repeated here.

[0081] (1) The calculation process of image structural feature similarity is as follows:

[0082] S201: Calculate the gradient magnitude map GM of the plaintext image P according to formulas (8) and (9) respectively. P Gradient magnitude map GM of encrypted image E E ;

[0083]

[0084]

[0085] Where "*" represents the linear convolution operator, M k k = 1, 2, 3, 4 represents the gradient operators in four directions, such as Figure 3 As shown, the weighting coefficients of the gradient operator decrease as the distance from the center pixel increases. The gradient operator is used to calculate the maximum weighted average value for each pixel. P(i,j) represents the pixel at the i-th row and j-th column position in the plaintext image P, E(i,j) represents the pixel at the i-th row and j-th column position in the encrypted image E, and mean2 represents the average value function.

[0086] Specifically, formulas (8) and (9) above take the plaintext image P and the ciphertext image E as input images, and utilize... Figure 3 The gradient operators in four directions perform convolution operations on the input image, and the maximum weighted average value in the four directions is used to represent the global structural features of the input image.

[0087] S202: first, the structural similarity map S between the gradient amplitude map of the plaintext image P and the gradient amplitude map of the encrypted image E is calculated according to formula (10) GM (i,j) is obtained; then the structural similarity map S GM (i,j) is averaged to obtain the image structure feature similarity S T ;

[0088]

[0089] Wherein, R is a normal number to avoid instability when the denominator converges to zero.

[0090] (2) The calculation process of the image content-aware feature similarity is as follows:

[0091] S301: the foreground segmentation threshold Th of the plaintext image P is calculated according to formula (11) by using the maximum inter-class variance method;

[0092] Th = p1(m1-mG) 2 +p2(m2-mG) 2 (11)

[0093] Wherein, m1 is the average gray value of the foreground, p1 is the ratio of the foreground pixel points to all pixel points of the image, m2 is the average gray value of the background, p2 is the ratio of the background pixel points to all pixel points of the image, and mG is the global gray mean value of the image;

[0094] S302: since the human visual system is more sensitive to the foreground content of the image, and the foreground of the image contains a large amount of image content, while the background has little effect on the understanding of the image content, the foreground segmentation threshold Th is used to filter the background matching pairs in the feature point set K3, and the number of remaining matching pairs is recorded as NOM; the foreground segmentation threshold Th is used to filter the background matching pairs in the feature point set K1, and the number of remaining matching pairs is recorded as NOP, and the image content-aware feature similarity S match ;

[0095]

[0096] The visual security evaluation method of the selectively encrypted image provided in the embodiment of the application uses a multi-directional gradient operator to obtain the structure of the plaintext image and the encrypted image, captures more accurate distortion of the edge area of the image, and calculates the average value of the structural similarity map of the image as the global structure feature similarity between images; meanwhile, considering the visual characteristics of the human eye, the background matching feature points are removed when calculating the content-aware feature similarity by using the maximum inter-class variance method, and the ratio of the number of key point matches to the number of foreground feature points of the original image is finally obtained as the content-aware feature similarity of the image.

[0097] Embodiment 3

[0098] On the basis of the above embodiments, the visual security evaluation method for selectively encrypted images provided by the embodiments of the present application provides a training method of a support vector regression model, can obtain a more accurate image visual security evaluation model, and other steps can refer to Embodiment 1 and / or Embodiment 2, which will not be repeated here.

[0099] S401: preparing a training data set, specifically comprising: collecting n clear text images, and using different types of selective encryption algorithms to encrypt the n clear text images to obtain corresponding n encrypted images; wherein one clear text image and its corresponding encrypted image constitute an image pair, and n image pairs constitute a training data set;

[0100] S402: respectively extracting image content perception feature similarity S match and image structure feature similarity S T of each image pair, and composing a feature similarity vector [S T ,S match ], thereby obtaining a feature similarity vector set H = {H1, H2, …, H n} of the training data set, wherein H n represents the feature similarity vector of the nth image pair.

[0101] S403: collecting the mean opinion score MOS (Mean Opinion Score) of all image pairs, and constructing to obtain a set S = {S1, S2, …, S n} wherein S n represents the mean opinion score MOS of the nth image pair.

[0102] S404: according to the set H and the set S, using formula (13) to establish the implicit relationship between each feature similarity vector and the corresponding mean opinion score MOS, and obtaining an image visual security evaluation model Pr_Model;

[0103] Pr_Model = SVR train(H,S) (13)

[0104] Wherein, SVR represents a support vector regression model.

[0105] Specifically, support vector regression (Support Vector Regression, SVR) is proposed by a Russian mathematician Vapnik, which can better solve the problem of limited sample and local minimum value, has strong generalization ability, can prevent overfitting to a certain extent, has strong robustness, and the basic idea behind it is to find the best fitting curve.

[0106] In this embodiment of the invention, the principle of using a support vector regression model to predict the visual security evaluation result of an image is as follows: In the training dataset D = {(H1,S1),(H2,S2),…,(H... n ,S n )}(H i ∈R m ,S i Learning on (∈R, m=1,2,3,…) yields a regression model such that f(ω,H) i ) and S i The difference between them should be as small as possible, f(ω,H) i The formula is as follows:

[0107] f(ω,H i )=ω T Φ(H i )+b (14)

[0108] Among them, H i S is the input feature similarity vector. i This is the mean opinion score (MOS) of the image, where ω is the weight vector, T is the transpose of the vector, b is the bias constant, and Φ(H) is the mean opinion score. i The kernel function (RBF) transforms nonlinear input-output problems into high-dimensional solutions for nonlinear regression problems. In this embodiment, the RBF kernel is used, and its formula is as follows:

[0109]

[0110] To verify the performance of the trained image visual security evaluation model, testing the model is also included, specifically: calculating the feature similarity vector H of a pair of test images. test Then, input it into the image visual security assessment model to obtain the corresponding visual security assessment prediction result S. test The prediction process can be expressed as formula (16):

[0111]

[0112] If S test If the MOS (Mean Opinion Score) of the test image is close to the true mean opinion score, then the performance is good and it can be applied.

[0113] Example 4

[0114] To verify the effectiveness of the method of the present invention, the present invention also provides the following experiments.

[0115] The experimental simulation platform is as follows: CPU: Intel(R) Core(TM) i5-7200, 2.50 GHz; Memory: 8.00 GB; Operating system: Windows 10; Coding tool: Matlab 2019b. The data set used in the experiment of the method is the public IVC-selectEncrypt and PEID. The IVC-SelectEncrypt database is composed of 8 original images, 200 encrypted images of which are generated using 5 different encryption algorithms with 5 different encryption degrees, and the subjective score MOS of each image in the database is in the range of [1, 5]. The PEID database has 1080 encrypted images, which are encrypted by 20 original images using 10 encryption schemes, and the subjective score of each image in the database is in the range of [0, 6]. In this experiment, features are extracted from the plaintext image and the corresponding encrypted image, and the similarity of the features is calculated, then the support vector regression is used to take the feature similarity and the MOS value of the encrypted image as training data to obtain the image visual security evaluation model, and finally the visual security score of the ciphertext image is predicted by the image visual security evaluation model.

[0116] (I) Comparative analysis

[0117] To demonstrate the performance of the VSI we proposed, we tested it on two benchmark image databases and compared it with nine state-of-the-art VSIs in the following three aspects, including PSNR, SSIM, VIF, LSS, ESS, LFBVS, and VSI-Canny.

[0118] (1) Confidence analysis

[0119] Confidence is used to determine the degree to which the visual security index actually reflects the human judgment result. Let D be an image database, where each image I has a subjective score s I and an objective score o I calculated by the visual security index. Our goal is to evaluate the consistency between the subjective score s I and the corresponding objective score o I . For the subjective score x and the objective score, the confidence C x is defined as:

[0120] C x = V max (x) - V min (x) (17)

[0121] where V max (x) and V min (x) are defined as follows:

[0122]

[0123]

[0124] V max (x) represents the maximum objective score of images in the database whose subjective scores do not exceed x. Similarly, V min (x) represents the minimum objective score of images in the database whose subjective score is no less than x. Because different VSI metrics have different output ranges, the confidence level C is set to a different value to ensure comparability between the different metrics. x Standardized to C′ x Based on C x Normalized confidence C′ x Normalized average confidence level μ D Normalized standard deviation σ D and normalized maximum confidence sum D The calculation formula is as follows:

[0125]

[0126]

[0127]

[0128]

[0129] Figure 4 The confidence evaluation results on the IVC-SelectEncrypt database are described, where MOS and the corresponding VSI objective score are plotted as black dots, and the nonlinear fitting results of MOS and VSI objective scores are plotted as red dashed lines. max (x), V min (x) and C x These are drawn as solid blue, green, and black lines, respectively. A better VSI should have a lower C. x Value, and C x The changes should be stable. From Figure 4 It can be clearly seen that for several VSIs such as SSIM and LFBVS, as the MOS decreases, C... x The C values ​​of VIF, ESS, and LSS show a clear increasing trend. x Exhibiting significant instability, PSNR, VSI-Canny, and our proposed VSI are more stable and have lower C values ​​than other metrics. x value.

[0130] Table 1 lists the confidence evaluation results of all VSI on PEID and IVC-SelectEncrypt databases, and the best value in evaluation indicators is marked in bold, and the worst indicator score is marked with an underline. It can be seen from Table 1 that the VSI proposed by us is the best overall compared with other compared VSI. Specifically, the best values are obtained on μ D and max D on PEID database and σ D and max D on IVC-SelectEncrypt database, in addition, although the VSI proposed by us is not the best on σ D on PEID database and μ D on IVC-SelectEncrypt database, but it is very close to the best value. From these observations, it can be seen that the overall performance of the VSI proposed by us is significantly better than all the compared VSI.

[0131] Table 1 Confidence evaluation results of VSI on PEID and IVC-SelectEncrypt databases

[0132]

[0133] (2) Monotonicity analysis

[0134] Monotonicity is used to determine the correlation between MOS and VSI scores, and two standard Spearman rank correlation coefficient (SRCC) and Kendall rank correlation coefficient (KRCC) are used to determine the monotonicity between objective scores and subjective perception. The calculation formulas of SRCC and KRCC are as follows:

[0135]

[0136]

[0137] Where N represents the number of samples, d i represents the difference between the subjective score ranking of the i-th image and the objective score ranking. N c is the number of consistent pairs in the data set, that is, the number of data levels in the MOS set and the VSI score set that are the same, N dis the number of inconsistencies in the dataset, i.e. the number of times the order of the data levels in the MOS set and the VSI score set are not the same. The larger the SRCC and KRCC values are, the higher the monotonicity between the subjective scores and the objective scores is. From Table 2 we can see that the VSI we proposed has the best SRCC and KRCC on both databases, which means that our VSI has a strong correlation with the subjective perception.

[0138] Table 2 Monotonicity evaluation results of VSI on PEID and IVC-SelectEncrypt databases

[0139]

[0140] (3) Accuracy analysis

[0141] The Person linear correlation coefficient (PLCC) and the root mean squared error (RMSE) are used to evaluate the accuracy between the MOS and the VSI objective scores. Since the range of the MOS values of different database images is different, for example, IVC-SelectEncrypt is [1, 5] and PEID is [0, 6], the range of the objective results of different evaluation indicators is also different, so before calculating the PLCC and the RMSE, we use the five-parameter nonlinear fitting function of equation (26) to nonlinearly fit the objective scores and the subjective scores, and map the objective scores to the subjective scores.

[0142]

[0143] where x is the visual security score calculated by the VSI, f(x) is the result after nonlinear fitting, τ i (i = 1, 2, 3, 4, 5) are the parameters of the function f(x). A good VSI is expected to obtain larger SRCC, KRCC and PLCC values, and smaller μ D , σ D , max D and RMSE values.

[0144]

[0145]

[0146] s i and p i represent the MOS score and the VSI score of the ith image, respectively, and MOS score average and VSI score average, respectively. From Table 3 we can see that our proposed VSI has the best PLCC and RMSE on both databases, which indicates that our VSI has a strong prediction accuracy.

[0147] Table 3. Accuracy evaluation results of VSI on PEID and IVC-SelectEncrypt databases

[0148]

[0149] (B) Evaluation on different quality ranges

[0150] The selectively encrypted images usually have low and medium visual quality. Therefore, in order to evaluate the performance of these VSI indicators more comprehensively, we evaluate the performance of these VSI in different quality ranges (i.e., low, medium, high). We divide the images of the IVC-SelectEncrypt database into three different quality ranges according to the MOS scores, and evaluate the performance of all related VSI in different quality ranges. The detailed division information can be found in Table 4. From Table 4 we can observe that in the low quality and medium quality ranges, our proposed indicators show more stable and superior performance than all other VSI, which is matched for the visual security evaluation task of selectively encrypted images with lower visual quality. In contrast, other indicators, such as SSIM, show satisfactory performance in the high quality range, but cannot provide better results in the low quality range.

[0151] Table 4. Overall performance comparison in different ranges

[0152]

[0153]

[0154] (C) Evaluation on different types of encryption or distortion

[0155] We also evaluate the evaluation effect of all indicators on different encryption types on the IVC-SelectEncrypt database to evaluate the performance of all indicators more comprehensively. Table 5 and Table 6 report the performance results of all encryption types that appear in the test database, respectively.

[0156] As can be seen in Table 5 and Table 6, our proposed metric achieves the best hit number 7 times in monotonicity and accuracy compared to other metrics, and performs best on the trad, trunc and iwind_nec encryption types on the IVC-SelectEncrypt database, and in addition, our metric performs very close to the best in other encryption types where it does not achieve the best. Thus, we draw the following conclusions from these experimental results. First, no metric achieves the best on all encryption types. Second, our proposed VSI outperforms the rest of the 7 metrics in overall performance on most encryption types.

[0157] Table 5 Overall performance comparison of different encryption types on the IVC-SelectEncrypt database (SRCC)

[0158]

[0159] Table 6 Overall performance comparison of different encryption types on the IVC-SelectEncrypt database (PLCC)

[0160]

[0161] Finally, it should be noted that the above examples are only used to illustrate the technical solutions of the present application, and are not intended to limit the same; although the present application has been described in detail with reference to the foregoing examples, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing examples, or make equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method of visual security evaluation of selectively encrypted images, characterized in that, The method comprises the following steps: Step 1: respectively detecting the plaintext image P and the corresponding encrypted image E to extract key points in the images, and respectively calculating the feature vectors of each key point in the plaintext image P and the encrypted image E, so as to obtain a feature point set K1 of the plaintext image P and a feature point set K2 of the encrypted image E; Step 2: respectively calculating the gradient amplitude graph of the plaintext image P and the gradient amplitude graph of the encrypted image E; Step 3: performing feature matching on the feature point set K1 and the feature point set K2 by using the minimum Euclidean distance, and recording the matched feature point set as K3; Step 4: calculate the structural similarity map between the gradient magnitude map of the plaintext image P and the gradient magnitude map of the encrypted image E, and calculate the image structural feature similarity S between the plaintext image P and the encrypted image E according to the structural similarity map T ; Step 5: Calculate the image content-aware feature similarity S between the plaintext image P and the encrypted image E according to the feature point sets K1 and K3 match ; Step 6: constructing and training a support vector regression model to obtain an image visual security evaluation model; Step 7: input the feature similarity vectors [S T ,S match ] of the plaintext image P and the encrypted image E to the trained image visual security evaluation model to obtain the image visual security evaluation result.

2. A method of visual security evaluation of a selectively encrypted image according to claim 1, characterized in that, In step 1, the SIFT algorithm is used as a feature point detection method to detect key points in the image and calculate the feature vectors of the key points.

3. A method of visual security evaluation of a selectively encrypted image according to claim 1, characterized in that, Step 2 specifically comprises: The gradient magnitude map GM of the plaintext image P is calculated according to formula (8) and formula (9) respectively P and the gradient magnitude map GM of the encrypted image E E ; where "*" denotes a linear convolution operator, M k , k = 1, 2, 3, 4 denote the gradient operators in four directions, P(i, j) denotes the pixel at the i-th row and j-th column position in the plaintext image P, E(i, j) denotes the pixel at the i-th row and j-th column position in the encrypted image E, and mean2 denotes the mean function.

4. A method of visual security evaluation of a selectively encrypted image according to claim 3, characterized in that, Step 4 specifically comprises: First, the structural similarity map S between the gradient amplitude map of the plaintext image P and the gradient amplitude map of the encrypted image E is calculated according to formula (10) GM (i,j); and then the structural similarity map S GM (i,j) is averaged to obtain the image structural feature similarity S T ; Wherein, R is a normal number to avoid instability when the denominator converges to zero.

5. A method of visual security evaluation of a selectively encrypted image according to claim 1, characterized in that, Step 5 specifically comprises: Step 5.1: calculating the foreground segmentation threshold Th of the plaintext image P according to formula (11) by using the maximum inter-class variance method; Th = p1(m1 - mG) 2 + p2(m2 - mG) 2 (11) Wherein, m1 is the average gray value of the foreground, p1 is the ratio of the foreground pixel points to all pixel points of the image, m2 is the average gray value of the background, p2 is the ratio of the background pixel points to all pixel points of the image, and mG is the global gray mean value of the image; Step 5.2: Filter the background matching pairs in the feature point set K3 using the foreground segmentation threshold Th, and record the number of remaining matching pairs as NOM; filter the background matching pairs in the feature point set K1 using the foreground segmentation threshold Th, and record the number of remaining matching pairs as NOP, and calculate the image content-aware feature similarity S according to formula (12) match ; 6. A method of visual security evaluation of a selectively encrypted image according to claim 1, characterized in that, Step 6 specifically comprises: Step 6.1: preparing a training data set, specifically including: collecting n plaintext images, and encrypting the n plaintext images by using different types of selective encryption algorithms to obtain corresponding n encrypted images; wherein, one plaintext image and its corresponding encrypted image constitute an image pair, and n image pairs constitute the training data set; Step 6.2: Extract the image content feature similarity S of each image pair according to step 4 and step 5 respectively match and the image structure feature similarity S T , and form a feature similarity vector [S T , S match ], so as to obtain the feature similarity vector set H = {H1, H2, …, H n} of the training data set, H n represents the feature similarity vector of the nth image pair; Step 6.3: Collect the mean opinion scores MOS for all image pairs and construct the resulting set S = {S1, S2,..., S n}, S n denotes the mean opinion score MOS for the n-th image pair. Step 6.4: according to the set H and the set S, an implicit relationship between each feature similarity vector and the corresponding mean opinion score MOS is established by using formula (13), and an image visual security evaluation model Pr_Model is obtained; Pr_Model = SVR train ( H,S ) (13) Wherein, SVR represents a support vector regression model, and a radial basis function kernel is used as a kernel function.

Images