Perceptual hash method for image visual security evaluation
By extracting multi-scale features of encrypted images through quaternion discrete wavelet transform, the problem of inaccurate visual security assessment of encrypted images in existing technologies is solved, and efficient and accurate visual security assessment of images is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-25
- Publication Date
- 2026-03-13
AI Technical Summary
Existing technologies struggle to accurately assess the visual security of encrypted images, especially in low-quality images. Traditional methods fail to effectively consider the perceptual characteristics of the human visual system, and hash algorithms are overly sensitive to subtle changes in images, leading to misjudgments and inaccurate assessments.
Quaternion discrete wavelet transform is used to extract multi-scale low-frequency brightness features, texture features and image complexity features from directional complementary subbands, construct hash values, and generate objective scores that are more in line with human subjective perception by comprehensively considering brightness, texture and complexity features.
It achieves efficient and accurate visual security assessment of encrypted images, effectively resisting image transformation and distortion, and providing an objective evaluation that is more in line with human subjective perception.
Smart Images
Figure CN121661364A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image visual security assessment technology, and in particular to a perceptual hashing method for image visual security assessment. Background Technology
[0002] With the development of the internet, images have been widely used due to their numerous advantages. However, along with the widespread use of images, visual information security issues have become increasingly serious, and the visual security assessment of encrypted images has become a key issue that urgently needs to be addressed. Existing research methods can be mainly divided into three categories. The first category is based on cryptographic statistical properties, using traditional indicators such as image histogram balance, information entropy value, and pixel correlation to construct an evaluation system. However, this type of method only focuses on the statistical information of encrypted images and does not consider the perceptual characteristics of the human visual system, making it difficult to accurately assess the visual security of perceived encrypted images. The second category uses image quality assessment (IQA) indicators, such as peak signal-to-noise ratio and structural similarity. Since image quality assessment is mainly designed for medium- to high-quality images, while perceived encrypted images are mostly low-quality images, and low-quality images may also contain highly sensitive visual information, this method cannot effectively assess the visual security of images. The third category is based on visual security assessment indicators (VSA) of the human visual system. Although this method attempts to quantify the amount of visual information leakage by comparing the similarity between plaintext and perceived encrypted image content, its accuracy still has room for improvement. Meanwhile, traditional hash algorithms, which rely solely on low-level pixel values or simple features (such as color histograms) to generate hash values, are overly sensitive to subtle changes in images. Slight adjustments to brightness, contrast, or the addition of noise can cause significant changes in hash values, leading to misidentification of similar images. Furthermore, they struggle to accurately identify image similarities when faced with complex image transformations, failing to meet the demand for precise image similarity assessment in encrypted image visual security evaluations.
[0003] Given the limitations of the aforementioned methods, perceptual image hashing technology exhibits unique advantages due to its ability to generate feature sequences from the visual content level and its sensitivity in measuring image distortion. By extracting image features to generate hash values, it can effectively resist geometric transformations and content distortion, providing a new direction for the visual security evaluation of encrypted images. Based on this, this invention proposes a perceptual hashing method for image visual security evaluation, which achieves more efficient and accurate visual security assessment of encrypted images by comprehensively considering low-frequency brightness features, texture features, and image complexity features. Summary of the Invention
[0004] For perceptually encrypted images, low quality does not necessarily equate to high security. Since humans are the ultimate recipients of images, subjective human evaluation is the most authoritative method for assessing the security of encrypted images. However, subjective evaluation is time-consuming and laborious. Therefore, there is an urgent need for a more effective objective evaluation metric to approximate subjective evaluation and achieve the same effect. To address the above problems, this invention proposes a perceptual hashing method for image visual security evaluation.
[0005] To achieve the above objectives, the present invention provides the following technical solution: A perceptual hashing method for image visual security evaluation employs quaternion discrete wavelet transform to extract multi-scale low-frequency brightness features, texture features, and image complexity features from directional complementary subbands. The extracted features are used to construct hash values, and the similarity of image hashes is measured to obtain an objective score between a reference image and a perceptually encrypted image. The method includes the following steps: Step 1), acquire the reference image and the perceptual encrypted image, and perform bicubic interpolation preprocessing on the reference image and the perceptual encrypted image; Step 2): The preprocessed reference image and the perceptually encrypted image are pixelated. Adjacent pixels are grouped and recombined to obtain three horizontally complementary sub-bands. , and three vertically complementary sub-bands And preserves the distortion introduced by the encryption operation to the image; Step 3) Extract multi-scale low-frequency brightness features, texture features, and image complexity features from the directional complementary subband using quaternion discrete wavelet transform. The directional information of the image can be represented by pure quaternions using quaternion discrete wavelet transform, with the horizontal axis being... , vertical , among them , , Representing the imaginary unit, the directional complementary subbands are used as the three imaginary parts of the quaternion. The low-frequency subband coefficients of the image are used to consider local brightness and extract brightness features. By selecting high-frequency subbands related to texture, wavelet coefficients of different frequencies are used to capture texture features of images at different scales. The complexity characteristics are calculated by combining the azimuth energy of coordinate points within a specific sub-band with the azimuth energy of neighboring coordinate points around these points. ; Step 4) The extracted brightness features, texture features, and image complexity features are concatenated together to construct two hash sequences for the original image and the perceptually encrypted image; Step 5), calculate the difference between the two hash sequences. Norms serve as a visual security score for perceptually encrypted images.
[0006] Specifically, in step 2), three adjacent pixels are grouped together, and the first, second, and third pixels of each group are recombined to obtain a horizontal complementary sub-band. Divide three adjacent pixels into a group, and recombine the first, second, and third pixels of each group to obtain vertical complementary sub-bands. .
[0007] Specifically, in step 3), the quaternion transform matrix of the original image P and the encrypted image E is obtained by using quaternion discrete wavelet transform. The low-frequency sub-band coefficients of the image are used to consider local brightness to extract brightness features. The mid-to-high frequency sub-bands are retained, and the multi-scale feature cross-scale fusion mechanism is considered. The texture features are calculated using the horizontal energy coefficient, vertical energy coefficient, and diagonal energy coefficient. The consistency of directional energy between the center point and the surrounding points in a specific sub-band is compared to calculate the complexity features.
[0008] Specifically, in step 4), the low-frequency brightness features, texture features, and pattern complexity features calculated from the quaternion discrete wavelet transform are quantized and processed according to... The formula shows how to construct a hash sequence by concatenating the elements together, where , , These are the weighting coefficients.
[0009] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects: (1) The present invention constructs complementary subbands, decomposing the input reference image and the perceptual encrypted image into complementary subbands in the horizontal and vertical directions. This process does not lose any information, and the distortion caused by the encryption operation to the image is preserved in the directional complementary subbands. Therefore, the features extracted from these subbands not only contain the semantic content information of the original image, but also incorporate the structural perturbation introduced by the encryption operation, which can distinguish images with different content.
[0010] (2) This invention is oriented towards the fusion of multi-dimensional features of human visual perception. Since the human visual system is extremely sensitive to the degradation of brightness, texture and pattern features, quaternion discrete wavelet transform is used to consider the structural features and content perception features of the image. Pure quaternions are used to represent the image orientation information. Brightness features, texture features and complexity features are extracted through different frequency sub-band coefficients. This deep fusion of multi-dimensional features enables the final generated image hash to fully reflect the visual content of the image, thereby providing an objective score that is more in line with the subjective perception of the human eye.
[0011] (3) This invention comprehensively considers image brightness features, texture features, and complexity features, and uses quaternion discrete wavelet transform to decompose the image into a multi-scale space. Based on this, it utilizes... Norms construct hash sequences, enhancing the discriminative power of features and thus ensuring the accuracy of the evaluation. Attached Figure Description
[0012] Figure 1 This is a flowchart of the perceptual hashing method for image visual security evaluation of the present invention; Figure 2 This is a schematic diagram of dividing complementary subbands; Figure 3 This is a schematic diagram of the Q-channel subband after quaternion discrete wavelet transform and a diagram of the selected feature extraction subband. Detailed Implementation
[0013] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings: like Figure 1 As shown, the perceptual hashing method for image visual security evaluation according to the present invention includes the following steps: Step 1) Obtain the reference image and the perceptually encrypted image, and preprocess them. Use bicubic interpolation to convert the input image size to T×T, and use Gaussian low-pass filtering to eliminate noise.
[0014] Specifically, a new pixel value is generated by calculating the weighted average of the 16 neighboring pixels surrounding the target pixel, and the image is scaled to the target size T×T. A Gaussian low-pass filter is applied to remove noise and smooth the image; the method used provides high-quality input for subsequent image hashing calculations and applications.
[0015] In step 2), the input reference image and the perceptually encrypted image are grouped into adjacent pixels and recombined to obtain complementary subbands in the horizontal and vertical directions, while retaining the distortion introduced by the encryption operation to the image.
[0016] Specifically, such as Figure 2 A multi-resolution representation is established to simulate the directionality and hierarchy of the human visual system. In practice, three horizontal complementary sub-bands and three vertical complementary sub-bands are obtained. Three adjacent pixels are grouped together, and the first, second, and third pixels of each group are recombined to obtain the horizontal complementary sub-bands. Divide the left and right three adjacent pixels into a group, and recombine the 1st, 2nd, and 3rd pixels of each group to obtain vertical complementary sub-bands. A quaternion model is proposed, in which the orientation information of an image can be represented by pure quaternions. as well as The quaternion transform matrix for the original image P and the encrypted image E is obtained using quaternion discrete wavelet transform, and the formula is as follows: Furthermore, the quaternion discrete wavelet transform and the discrete wavelet transform (DWT) process can be connected by a unit quaternion parameter µ. In the inverse process of the quaternion discrete wavelet transform, the final result must be a pure quaternion matrix; that is... .like Figure 3 As shown, to minimize energy loss, the Q portion is chosen as the optimal feature extraction channel. The low-frequency subband of Q is used... Indicated. The high-frequency subband of Q is used... The expression represents the decomposition level, where σ represents the number of decomposition levels and L represents the directionality of the transformed subband. .
[0017] Step 3) Use quaternion discrete wavelet transform to extract multi-scale low-frequency brightness features, texture features and image complexity features from directional complementary subbands.
[0018] First, extracting multi-scale low-frequency brightness features is crucial because the human eye is highly sensitive to changes in image brightness, and low-frequency sub-bands can effectively represent the brightness distribution of an image. For a normal image P and its perceptually encrypted image E, the corresponding low-frequency sub-band brightness features can be calculated as follows: Secondly, texture features are extracted. High-frequency sub-bands relevant to texture are selected from the output of the quaternion discrete wavelet transform. Wavelet transform captures local features of the image at different scales by decomposing the image into wavelet coefficients of different frequencies. Low-frequency sub-bands are removed according to different transform decomposition levels, while mid- and high-frequency sub-bands are retained. Feature maps based on different directions of the Q-channel are generated through special processing. Under the first-level decomposition, the channels... Only keep and Subbands are processed, and feature maps are generated in the H, V, and D directions through the inverse process of quaternion discrete wavelet transform. In the second-level decomposition, while retaining and Based on the sub-bands, the feature mapping generation operation from the first-level decomposition is performed to generate feature maps. Finally, for the low-frequency sub-bands... Two-layer channel Downsampling is performed to obtain a new sub-band third-layer channel. Calculate the average coefficient matrix of each sub-block. Considering the multi-scale feature cross-scale fusion mechanism, the horizontal energy coefficient, vertical energy coefficient, and diagonal energy coefficient are expressed as follows: and Texture features are calculated as follows: Finally, image complexity features are extracted. Complexity features are calculated by combining the azimuth energy of coordinate points within a specific sub-band with the azimuth energy of neighboring coordinate points around these points. The maximum energy coefficient is then calculated. By calculating the directional energy of the subband coefficients in a 3×3 local region, local features of the image are extracted, and the mode complexity of this region is characterized as follows: The formula for pattern complexity is: In step 4), the extracted brightness features, texture features, and image complexity features are concatenated to construct two hash sequences.
[0019] Specifically, the initial hash sequence is constructed as follows: . Let i be the i-th element in the initial hash sequence. To reduce storage space, we will... Quantized to integers The quantification formula is: The final hash sequence L is a sequence of d integers: Step 5) Utilize The norm is used to measure the similarity of image hashes, resulting in an objective score between the reference image and the perceptually encrypted image.
[0020] Specifically, let's set and These are the hash sequences of the original image and the encrypted image, respectively. The two hash sequences... Norm is defined as: If S is less than a set threshold T, the reference image is considered similar to the perceptually encrypted image. A smaller S indicates more similar images to the input hash sequence. Therefore, for different images, their corresponding S should be a larger value.
[0021] The experimental results of this invention are as follows: The performance of this invention was evaluated in terms of confidence, monotonicity, linearity, and accuracy by comparing it with other IQA methods and VSA. Tables 1 and 2 list the overall performance metrics of all IQA and VSA methods on the IVC-SelectEncrypt and PEID databases, respectively, with the best-performing metrics highlighted in bold. In the IVC-SelectEncrypt database, this invention exhibits the best performance. , PLCC, SRCC, KRCC, and RMSE values. It has the best values in the PEID database. PLCC, SRCC, and KRCC values. Clearly, the performance of this invention using IVC-SelectEncrypt and PEID outperforms other methods. More than half of the metrics show optimization.
[0022] Table 1 Overall performance of the IVC-SelectEncrypt database Table 2 Overall performance of the PEID database
Claims
1. A perceptual hashing method for image visual security evaluation, which employs quaternion discrete wavelet transform to extract multi-scale low-frequency brightness features, texture features, and image complexity features from directional complementary subbands, constructs hash values from the extracted features, measures the similarity of image hashes, and obtains an objective score between a reference image and a perceptually encrypted image. The method includes the following steps: Step 1), acquire the reference image and the perceptual encrypted image, and perform bicubic interpolation preprocessing on the reference image and the perceptual encrypted image; Step 2): The preprocessed reference image and the perceptually encrypted image are pixelated. Adjacent pixels are grouped and recombined to obtain three horizontally complementary sub-bands. , and three vertically complementary sub-bands And preserves the distortion introduced by the encryption operation to the image; Step 3) Extract multi-scale low-frequency brightness features, texture features, and image complexity features from the directional complementary subband using quaternion discrete wavelet transform. The directional information of the image can be represented by pure quaternions using quaternion discrete wavelet transform, with the horizontal axis being... , vertical , among them , , Representing the imaginary unit, the directional complementary subbands are used as the three imaginary parts of the quaternion. The low-frequency subband coefficients of the image are used to consider local brightness and extract brightness features. By selecting high-frequency subbands related to texture, wavelet coefficients of different frequencies are used to capture texture features of images at different scales. The complexity characteristics are calculated by combining the azimuth energy of coordinate points within a specific sub-band with the azimuth energy of neighboring coordinate points around these points. ; Step 4) The extracted brightness features, texture features, and image complexity features are concatenated together to construct two hash sequences for the original image and the perceptually encrypted image; Step 5), calculate the difference between the two hash sequences. Norms serve as a visual security score for perceptually encrypted images.
2. The perceptual hashing method for image visual security evaluation as described in claim 1, characterized in that: In step 2), three adjacent pixels are grouped together, and the first, second, and third pixels of each group are recombined to obtain a horizontal complementary sub-band. Divide three adjacent pixels into a group, and recombine the first, second, and third pixels of each group to obtain vertical complementary sub-bands. .
3. A perceptual hashing method for image visual security evaluation, characterized in that: In step 3), the quaternion discrete wavelet transform is used to obtain the quaternion transform matrix of the original image P and the encrypted image E, and the low-frequency subband coefficients of the image are used to consider the local brightness to extract brightness features. The mid-to-high frequency sub-bands are retained, and a multi-scale feature cross-scale fusion mechanism is considered. Texture features are calculated using horizontal energy coefficients, vertical energy coefficients, and diagonal energy coefficients. The consistency of directional energy between the center point and surrounding points within a specific sub-band is compared to calculate complexity features.
4. A perceptual hashing method for image visual security evaluation, characterized in that: In step 4), the low-frequency brightness features, texture features, and pattern complexity features calculated from the quaternion discrete wavelet transform are quantized and then... The formula shows how to construct a hash sequence by concatenating the elements together, where , , These are the weighting coefficients.