Image retrieval method and system based on improved feature clustering

Abstract

The invention discloses an image retrieval method and system based on improved feature clustering, and the method comprises the steps: carrying out the clustering operation processing and encryption processing of an image data set, and obtaining a hierarchical structure and an encrypted image of the image data set; encrypting the hierarchical structure of the image data set through an enhanced feature encryption algorithm to obtain a searchable index tree; sequentially performing image feature extraction and encryption processing on the to-be-retrieved image data to obtain a searchable trap door; and taking the Euclidean distance as a retrieval reference, calculating the searchable trap door and the searchable index tree, obtaining similar images in the image data set, and decrypting the encrypted images to obtain the similar images in the image data set. The method can focus on related categories during retrieval, so that the clustering quality is ensured, and the image retrieval efficiency is improved. The image retrieval method and system based on improved feature clustering can be widely applied to the technical field of data retrieval.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data retrieval, in particular to an image retrieval method and system based on improved feature clustering. BACKGROUND

[0002] With the popularity of intelligent terminal devices and the rapid development of Internet of Things technology, images as a kind of core multimedia data are being generated and shared at an unprecedented speed in Internet of Things application scenarios such as intelligent medical treatment, intelligent transportation, and social networks. According to statistics, more than 3.2 billion images are shared every day worldwide. The massive image data poses a huge challenge to the storage and computing capacity of terminal devices. Therefore, outsourcing data to cloud servers with strong resources has become a promising solution.

[0003] However, image data often contains a large amount of sensitive information (such as personal identity, medical records), and directly outsourcing it to the cloud server in plaintext form will cause serious privacy leakage risks. Encryption is an effective means to protect data confidentiality, but traditional encryption will make data unreadable, resulting in the failure of content-based image retrieval. Therefore, scholars have proposed privacy-protected ciphertext image retrieval technology, which allows users to directly search for similarity of encrypted images without decryption. Although existing technologies have made preliminary progress in this direction, they still face many serious challenges in actual deployment, especially in Internet of Things environments: first, retrieval accuracy and efficiency are difficult to achieve. On the one hand, many schemes rely on traditional global features or local features, which have limited feature representation capabilities, resulting in low retrieval accuracy. On the other hand, the use of local sensitive hashing and other approximate methods to improve efficiency will introduce errors, further reducing accuracy. At the same time, using linear index structures for similarity retrieval requires traversing the entire data set, which is inefficient and cannot meet the real-time needs of terminal devices. Second, there is a contradiction between security and performance. Some schemes that use homomorphic encryption or secure multi-party computation provide strong security guarantees, but their huge computational overhead and ciphertext expansion make it difficult to effectively utilize them in resource-constrained environments. In addition, commonly used kNN feature encryption algorithms are relatively efficient, but have been proven to be vulnerable to linear analysis attacks, with security vulnerabilities. While some improved schemes based on fault-tolerant learning enhance security, the noise introduced interferes with similarity calculation, compromising retrieval accuracy. SUMMARY

[0004] To solve the above technical problems, the purpose of the present application is to provide an image retrieval method and system based on improved feature clustering, which can focus on relevant categories during retrieval, ensuring clustering quality and improving image retrieval efficiency.

[0005] The first technical solution adopted by the application is: an image retrieval method based on improved feature clustering, comprising the following steps: The image data set is respectively subjected to clustering operation processing and encryption processing to obtain the hierarchical structure of the image data set and encrypted images; The hierarchical structure of the image data set is subjected to encryption processing through an enhanced feature encryption algorithm to obtain a searchable index tree; The image data to be retrieved is sequentially subjected to image feature extraction and encryption processing to obtain a searchable trapdoor; The Euclidean distance is taken as a retrieval benchmark to calculate the searchable trapdoor and the searchable index tree, obtain similar images in the image data set, and decrypt the encrypted images to obtain similar images in the image data set.

[0006] Further, the step of respectively subjecting the image data set to clustering operation processing and encryption processing to obtain the hierarchical structure of the image data set and encrypted images specifically comprises: The image data set is subjected to image feature extraction through a convolutional neural network model Regnet to obtain an image feature vector set; According to the similarity of image features, the image feature vector set is subjected to clustering operation through an optimized clustering algorithm to obtain the hierarchical structure of the image data set; The image data set is subjected to encryption through an AES algorithm to obtain encrypted images.

[0007] Further, the step of according to the similarity of image features, the image feature vector set is subjected to clustering operation through an optimized clustering algorithm to obtain the hierarchical structure of the image data set specifically comprises: The number of clustering centers is set and a control parameter is introduced, which is used to limit the maximum number of feature vectors in each cluster; According to the number of clustering centers, a plurality of vectors are selected as initial clustering center vectors, and the distance values of the image feature vectors and the initial clustering center vectors are calculated; The image feature vectors are divided into the nearest initial clustering center vectors according to the calculated distance values, until the entire image feature vector set is traversed, to obtain an initial divided image feature vector set, which includes a plurality of subspaces; The initial divided image feature vector set is screened according to the control parameter, and the subspaces with the number of feature vectors greater than a preset threshold are selected for redivision processing, until the number of feature vectors of all subspaces meets the control parameter, and the hierarchical structure of the image data set is output.

[0008] Further, the step of encrypting the hierarchy of the image data set through the enhanced feature encryption algorithm to obtain a searchable index tree specifically comprises: generating a key set according to the dimension of the feature vector in the hierarchy of the image data set; extending the feature vector in the hierarchy of the image data set to obtain an extended feature vector; encrypting the extended feature vector through the key set to obtain an encrypted feature vector; repeating the extension and encryption steps until all nodes in the hierarchy of the image data set are traversed, and a searchable index tree is constructed.

[0009] Further, the step of sequentially performing image feature extraction and encryption processing on the image data to be searched to obtain a searchable trapdoor specifically comprises: extracting image features of the image data to be searched through a convolutional neural network model Regnet to obtain an image feature vector to be searched; extending the image feature vector to be searched to obtain an extended image feature vector to be searched; encrypting the extended image feature vector to be searched through the key set to obtain a searchable trapdoor.

[0010] Further, the step of taking the Euclidean distance as a search benchmark, calculating the searchable trapdoor and the searchable index tree, obtaining similar images in the image data set, and decrypting the encrypted image to obtain similar images in the image data set specifically comprises: searching the root node of the searchable index tree, and at each level, calculating the inner product value between all node vectors of the level and the searchable trapdoor to obtain the Euclidean distance value; sorting the Euclidean distance values, selecting the node with the smallest Euclidean distance value as the path, and entering the next level until reaching the leaf node of the searchable index tree, and outputting the most similar encrypted image; based on the most similar encrypted image, decrypting the encrypted image through the AES algorithm to obtain similar images in the image data set.

[0011] Further, the calculation expression of the Euclidean distance value is specifically as follows:

[0012] In the above formula, represents the inner product value between all node vectors and the searchable trapdoor, represents the Euclidean distance value, represents a node vector of the searchable index tree, representing searchable trapdoors.

[0013] The second technical solution adopted by the present application is: an image retrieval system based on improved feature clustering, comprising: A first module is configured to perform clustering operation processing and encryption processing on the image data set respectively to obtain a hierarchical structure of the image data set and encrypted images. A second module is configured to perform encryption processing on the hierarchical structure of the image data set by using an enhanced feature encryption algorithm to obtain a searchable index tree. A third module is configured to perform image feature extraction and encryption processing on the image data to be retrieved in sequence to obtain a searchable trapdoor. A fourth module is configured to use Euclidean distance as a retrieval reference to calculate the searchable trapdoor and the searchable index tree, obtain similar images in the image data set, and decrypt the encrypted images to obtain similar images in the image data set.

[0014] The present application has the following advantages: the present application performs clustering operation processing and encryption processing on the image data set respectively to obtain a hierarchical structure of the image data set and encrypted images; the optimized clustering algorithm realizes the best balance between efficiency and accuracy through double parameter control, can quickly locate the target area along the optimal branch path, reduces the retrieval complexity from linear level to logarithmic level, maximizes the reduction of unnecessary computing overhead while ensuring the recall rate, and then performs encryption processing on the hierarchical structure of the image data set by using an enhanced feature encryption algorithm to obtain a searchable index tree; by adding random numbers, it is ensured that the encryption result is different each time for the same picture, preventing attackers from performing frequency analysis through repeated ciphertext, effectively protecting the association privacy of the data; further image feature extraction and encryption processing are performed on the image data to be retrieved in sequence to obtain a searchable trapdoor; the key visual content of the image can be effectively captured, thereby realizing fast and accurate similarity matching in image retrieval and other tasks; finally, Euclidean distance is used as a retrieval reference to calculate the searchable trapdoor and the searchable index tree, obtain similar images in the image data set, and decrypt the encrypted images to obtain similar images in the image data set, which can focus on related categories during retrieval, ensure the clustering quality, and improve the efficiency of image retrieval. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 is a step flowchart of an image retrieval method based on improved feature clustering of the present application; Figure 2 is a structural block diagram of an image retrieval system based on improved feature clustering of the present application; Figure 3This is a schematic diagram of the image retrieval method based on polynomial searchability provided in a specific embodiment of the present invention; Figure 4 This is a schematic diagram of the process of generating a searchable index tree provided in a specific embodiment of the present invention. Detailed Implementation

[0016] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. The step numbers in the following embodiments are only for ease of explanation and do not limit the order of the steps. The execution order of each step in the embodiments can be adapted according to the understanding of those skilled in the art.

[0017] First, it should be noted that the embodiments of the present invention first extract image features from the image dataset to obtain a set of image feature vectors. Based on the similarity of image features, an optimized clustering algorithm is used to cluster the set of image feature vectors to obtain a hierarchical structure. The image dataset is then encrypted using an image encryption algorithm to obtain encrypted images. Next, an enhanced feature encryption algorithm is used to encrypt the hierarchical structure to obtain a searchable index tree. Further, image features are extracted from the image data to be retrieved to obtain the feature vectors of the images to be retrieved. These feature vectors are then encrypted using the enhanced feature encryption algorithm to obtain searchable trapdoors. Finally, using Euclidean distance as the retrieval benchmark, the searchable trapdoors and the searchable index tree are calculated to return the most similar encrypted image. The returned encrypted image is then decrypted using an image decryption algorithm to obtain similar images from the image dataset. The optimized clustering algorithm flexibly adjusts the number of vectors in the clusters, focusing on relevant categories during retrieval, thus ensuring clustering quality and improving image retrieval efficiency. The invention reduces the error in similarity calculation caused by noise introduced by encryption through an enhanced feature encryption algorithm, ensuring the effectiveness of retrieval while achieving unlinkability of queries and resistance to linear attacks.

[0018] Reference Figure 1 This invention provides an image retrieval method based on improved feature clustering, which includes the following steps: S100. Perform clustering and encryption operations on the image data set to obtain the hierarchical structure of the image data set and the encrypted image. Specifically, the image dataset is processed by the Regnet convolutional neural network model to extract image features, resulting in a set of image feature vectors. Based on the similarity of the image features, the set of image feature vectors is clustered using an optimized clustering algorithm to obtain the hierarchical structure of the image dataset. Finally, the image dataset is encrypted using the AES algorithm to obtain encrypted images.

[0019] For the optimized clustering algorithm, the number of cluster centers is set and a control parameter is introduced. The control parameter is used to limit the maximum number of feature vectors contained in each cluster. Several vectors are selected as initial cluster center vectors according to the number of cluster centers, and the distance between the image feature vectors and each initial cluster center vector is calculated. The image feature vectors are divided to the nearest initial cluster center vector according to the calculated distance value, until the entire image feature vector set is traversed, resulting in an initially divided image feature vector set, which includes several subspaces. The initially divided image feature vector set is filtered according to the control parameter, and the subspaces with a feature vector number greater than a preset threshold are selected for re-division processing until the feature vector number of all subspaces meets the control parameter, and the hierarchical structure of the image data set is output.

[0020] In this embodiment, the RegNet convolutional neural network model is first used to extract image features from the image dataset, resulting in a set of image feature vectors. The RegNet neural network model improves performance by increasing network width and depth at different network layers, and the extracted features have strong discriminative power, capable of capturing subtle visual differences, thus returning more relevant search results.

[0021] Then, based on the similarity of image features, an optimized clustering algorithm is used to cluster the feature vectors of the images to obtain the hierarchical structure of the image data, as follows; Given two integer parameters and ,in This indicates the number of cluster centers, while This is used to limit the maximum number of feature vectors contained in each cluster. A common method for generating index trees using the K-means algorithm is to set... and Taking the same value would be a rigid approach. The value affects the clustering process of each branch, thus negatively impacting retrieval accuracy. Therefore, to improve the algorithm's adaptability in real-world scenarios and enhance retrieval accuracy, this invention proposes introducing a control parameter. This is designed to flexibly adjust the number of vectors in a cluster when building the index tree. Parameters The settings ensure that the number of nodes in each cluster is not too large at a finer granular level, which helps to improve the clustering quality and make the number of nodes in each cluster relatively balanced. The feature vector set is divided into groups using the K-means algorithm. The process in different clustering is as follows: First, randomly select... Each feature vector is used as the initial cluster center vector; then, the distance between each feature vector and each cluster center vector is calculated, and it is assigned to the nearest cluster. These two steps are performed iteratively until a preset number of iterations is reached or the cluster center vectors tend to stabilize; After the initial partitioning described above, the number of vectors in each subspace may still be too large. At this point, to ensure that the number of vectors in each subspace does not exceed a set value... The K-means algorithm will be executed again on the corresponding subspaces containing too many vectors. This partitioning process will continue until all subspaces meet the requirements. Finally, all feature vectors that are leaf nodes are clustered under a root node, thus constructing a hierarchical structure based on feature clustering, where each leaf node corresponds to a feature vector of an image. This algorithm uses parameters... By limiting the maximum capacity of each cluster and automatically triggering recursive clustering when the number of samples within a cluster exceeds a threshold, the system ensures a relatively balanced number of samples across subspaces. This dynamic adjustment mechanism significantly improves the similarity of samples within clusters, laying a solid foundation for efficient subsequent retrieval. Simultaneously, by establishing a hierarchical tree-like index, the system avoids traversing all data during queries, instead quickly locating the target region along the optimal branch path, reducing retrieval complexity from linear to logarithmic levels. This minimizes unnecessary computational overhead while maintaining high recall. Finally, the image data to be retrieved is encrypted using the AES algorithm to obtain an encrypted image. The pixel data stream formed after the input image file is read is obfuscated and diffused to completely randomize the image data, outputting a completely random ciphertext data stream unrelated to the original image. Only an authorized party with the same key can decrypt it using the AES decryption process to restore it to the original image. The AES algorithm has extremely high security, effectively resisting various attacks and ensuring the confidentiality of image content during transmission and storage.

[0022] S200. The hierarchical structure of the image data set is encrypted using an enhanced feature encryption algorithm to obtain a searchable index tree. Specifically, a key set is generated based on the dimension of the feature vectors in the hierarchical structure of the image dataset; the feature vectors in the hierarchical structure of the image dataset are expanded to obtain expanded feature vectors; the expanded feature vectors are encrypted using the key set to obtain encrypted feature vectors; the expansion and encryption steps are repeated until all nodes in the hierarchical structure of the image dataset are traversed to build a searchable index tree.

[0023] In this embodiment, firstly for the first Zhang Image eigenvectors Encryption is performed using an enhanced feature-based encryption algorithm, based on the feature vector. Dimensions Generate key set ,in, It is a random permutation sequence used to shuffle the positions of the elements within the feature vector. It is a matrix and its invertible matrix; further, for eigenvectors The feature vector is expanded to obtain the expanded feature vector. The expression is: ,in, It is an eigenvector The sum of the squares of all elements It is a random number; further processing of the expanded feature vector Using key sets Encryption is performed to obtain the encrypted feature vector. The expression is: .

[0024] The random permutation sequence aims to completely destroy the structure of the ciphertext. Even if an attacker can somehow deduce the meaning of a particular ciphertext component, they cannot determine which dimension that component corresponds to in the original feature vector. This significantly increases the difficulty of cryptanalysis, making it extremely difficult for attackers to deduce the encryption matrix from the ciphertext. Alternatively, understanding the semantic structure of feature vectors can effectively protect the patterns and distributions of the original features. Extending the feature vectors by adding random numbers ensures that each encryption result for the same image is different, preventing attackers from performing frequency analysis through repeated ciphertexts. More importantly, it achieves unlinkability of the ciphertexts, meaning the server cannot determine whether two different ciphertexts originate from the same original feature, thus effectively protecting the privacy of data association. Furthermore, based on the enhanced feature encryption algorithm described above, each node in the hierarchical structure is encrypted to obtain a searchable index tree. By uniformly applying the enhanced encryption algorithm to each node in the hierarchical structure, the entire index tree is transformed into a completely encrypted searchable structure. This means that when the cloud server performs a retrieval, every comparison and path selection from the root node to the leaf node is performed on encrypted text. It can only advance the search based on the encrypted comparison operations allowed by the encryption algorithm, and cannot know what the nodes it is comparing specifically represent, thus cutting off the chain of privacy leakage at the source.

[0025] S300: The image feature extraction and encryption processes are performed sequentially on the image data to be retrieved to obtain a searchable trapdoor; Specifically, the Regnet convolutional neural network model is used to extract image features from the image data to be retrieved, resulting in a feature vector of the image to be retrieved; the feature vector of the image to be retrieved is then expanded to obtain an expanded feature vector of the image to be retrieved; and the expanded feature vector of the image to be retrieved is then encrypted using a key set to obtain a searchable trapdoor.

[0026] In this embodiment, the RegNet convolutional neural network model is first used to extract image features from the image data to be retrieved, resulting in a feature vector. The feature vector generated by this neural network model is not only compact but also effectively captures the key visual content of the image, thus achieving fast and accurate similarity matching in tasks such as image retrieval. Then, an enhanced feature encryption algorithm is used to encrypt the feature vector of the image to be retrieved, obtaining a searchable trapdoor. The feature vector of the image to be retrieved is then... The feature vector of the image to be retrieved is then expanded. The expression is: ,in, It is a random number. It is a positive integer and is different in each query; then the expanded feature vector of the image to be retrieved... Using key sets Encryption is performed to obtain a searchable trapdoor. The expression is: ,in, yes The transpose of .

[0027] The expansion and encryption of the feature vector of the image to be retrieved is to construct a cryptographic trapdoor that can securely hide the query content while achieving accurate comparison in the ciphertext domain. The three independent random numbers introduced, especially the unique and positive integers for each query, are crucial. This ensures that even when repeatedly querying the same image, the generated trapdoors will be completely different each time. This randomization effectively prevents cloud servers from inferring user search intent or identifying duplicate queries by analyzing trapdoor patterns, thus protecting user search pattern privacy. Simultaneously, the random factor introduced into the extended vector... The random number used in hierarchical encryption By working together, necessary noise barriers are reduced in subsequent ciphertext similarity calculations. After algebraic operations, the Euclidean distance between the original feature vectors can be restored, thus achieving the same ranking result as plaintext retrieval without revealing any plaintext information.

[0028] S400. Using Euclidean distance as the retrieval benchmark, calculate the searchable trapdoors and searchable index trees to obtain similar images in the image dataset and decrypt the encrypted images to obtain similar images in the image dataset.

[0029] Specifically, the root node of the searchable index tree is searched. At each level, the inner product between all node vectors at that level and the searchable trapdoors is calculated to obtain the Euclidean distance value. The Euclidean distance values ​​are sorted, and the node with the smallest Euclidean distance value is selected as the path to enter the next level until the leaf node of the searchable index tree is reached, outputting the most similar encrypted image. Based on the most similar encrypted image, the encrypted image is decrypted using the AES algorithm to obtain similar images in the image dataset.

[0030] In this embodiment, Euclidean distance is used as the retrieval benchmark. The searchable trapdoors and the searchable index tree are compared to calculate the most similar encrypted image. The search begins from the root node of the searchable index tree. At each level, the inner product between all node vectors at that level and the searchable trapdoors is calculated. This value, in its encrypted state, represents the Euclidean distance, expressed as: ,in, It is a node vector of a searchable index tree. It is a searchable trapdoor. By transforming Euclidean distance into a dot product of encrypted vectors through algebraic transformations, it cleverly achieves efficient similarity calculation in the ciphertext domain. The calculation process does not require decrypting sensitive data, directly outputting a numerical result proportional to the plaintext distance while protecting query and index privacy, thus balancing retrieval accuracy and data security. The distance values ​​are then sorted, and the node with the smallest Euclidean distance (i.e., the most similar node) is selected as the path to the next level. This process is repeated until a leaf node is reached, indicating that the most similar encrypted image has been found. This sorting mechanism achieves efficient ciphertext retrieval by quantifying similarity. Minimizing the distance directly corresponds to the nearest encrypted image in the feature space, ensuring consistency between the retrieval result and the plaintext domain. Finally, the most similar encrypted image is decrypted using the AES algorithm to obtain similar images from the image dataset. This step achieves a balance between security and efficiency through symmetric decryption. The AES algorithm can quickly reconstruct the retrieved ciphertext image, ensuring that only authorized users can obtain the final plaintext result. This end-to-end encryption-partial decryption mechanism minimizes decryption overhead while maintaining overall data security, achieving a balance between privacy protection and practicality.

[0031] In summary, this invention extracts image features from an image dataset using a convolutional neural network model to obtain image feature vectors. Based on the similarity of image features, an optimized clustering algorithm is used to cluster these features, resulting in a structure. An enhanced feature encryption algorithm is then used to encrypt the index tree, yielding a searchable index tree structure. This invention applies an optimized clustering algorithm to flexibly adjust the number of vectors in each cluster, focusing on relevant categories during retrieval, thus ensuring clustering quality and improving image retrieval efficiency. Furthermore, the enhanced feature encryption algorithm reduces errors in similarity calculation caused by noise introduced by encryption, ensuring retrieval effectiveness while achieving query non-linkability and resistance to linear attacks.

[0032] Therefore, the embodiments of the present invention have the following advantages compared with the prior art: 1) This invention utilizes transfer learning and pre-trained deep learning models, leveraging convolutional neural networks pre-trained on large-scale image sets. These networks, through learning from massive amounts of data, have mastered general image patterns. After fine-tuning in this invention, they can deeply mine the semantic features of images, more accurately reflecting the true content of the images. Compared to traditional local and global feature retrieval methods, this invention significantly improves retrieval accuracy and better meets practical application needs.

[0033] 2) The optimized clustering algorithm used in this embodiment of the invention achieves an optimal balance between efficiency and accuracy through dual parameter control. Traditional methods use a single value to construct the index tree, resulting in uneven node distribution and affecting retrieval performance. By introducing two independent parameters—the number of clusters and the node capacity—recursive clustering is automatically triggered when the vectors within a cluster exceed the node capacity parameter, forming a structurally balanced index tree. This design reduces retrieval complexity from linear to logarithmic levels. The cloud server only needs to compare a fixed number of nodes at each layer to determine the optimal path, achieving rapid response. Simultaneously, the node capacity parameter ensures that each leaf node cluster maintains an appropriate size, avoiding feature confusion. Combined with an intelligent backtracking mechanism, extended searches are performed within relevant clustering domains, achieving retrieval accuracy comparable to linear retrieval. This dynamic balancing mechanism is suitable for real-world environments with complex data distribution, laying a solid foundation for efficient retrieval of multi-source image data.

[0034] 3) The enhanced feature encryption algorithm in this embodiment of the invention achieves a balance between security and practicality through a multi-randomization design. Based on the traditional secure kNN encryption algorithm, the algorithm introduces random permutations and multiple random numbers to asymmetrically expand the feature vector before encryption. This design ensures that the encryption result for the same image or query is completely different each time, providing strong unlinkability and effectively resisting linear analysis and known background attacks. Secondly, the carefully designed expansion structure ensures that after the ciphertext inner product operation, the random terms cancel each other out, and the result maintains a strictly linear relationship with the plaintext Euclidean distance, allowing the server to accurately compare similarities without knowing the plaintext. This algorithm provides strong security guarantees while supporting efficient ciphertext domain computation.

[0035] Finally, the embodiments of the present invention will be further explained and described with reference to the accompanying drawings: Specific Implementation Example 1: This embodiment is a fine-grained privacy-preserving image retrieval method based on searchable encryption, image retrieval, and cloud computing. First, the image owner extracts image features from the image dataset to obtain a set of image feature vectors. Based on the similarity of image features, an optimized clustering algorithm is used to cluster the image feature vector set, resulting in a hierarchical structure of the image dataset. The image dataset is then encrypted using the AES algorithm to obtain encrypted images, ensuring image security during transmission and retrieval. Next, the image owner uses an enhanced feature encryption algorithm to encrypt the hierarchical structure, resulting in a searchable index tree. Subsequently, the encrypted images and the searchable index tree are uploaded to a cloud server. The querying user extracts image features from the image data to be retrieved, obtaining feature vectors for the image to be retrieved. These feature vectors are then encrypted using the enhanced feature encryption algorithm to obtain searchable trapdoors, which are uploaded to the cloud server. The cloud server uses Euclidean distance as the retrieval benchmark to calculate the searchable trapdoors and the searchable index tree, obtaining a set of similar encrypted images in their encrypted state, which is then sent to the querying user. The querying user uses the decryption key sent by the image owner to decrypt similar encrypted images in their ciphertext state, obtaining the desired set of similar images. Notably, by calculating searchable trapdoors and a searchable index tree, the number of feature comparisons during retrieval is significantly reduced. Retrieval only requires filtering layer by layer along the branches of the searchable index tree to quickly locate the set of candidate images similar to the query features, avoiding the full comparison required by traditional linear retrieval. This provides crucial technical support for large-scale, high-dimensional feature-driven image retrieval tasks.

[0036] Detailed explanation is as follows: 1) The image owner extracts image features from the image dataset, obtaining a set of image feature vectors. Based on the similarity of image features, an optimized clustering algorithm is used to cluster the image feature vector set, resulting in a hierarchical structure of the image dataset. The image dataset is then encrypted using the AES algorithm to obtain an encrypted image. First, the image owner uses the RegNet convolutional neural network model to extract image features from the image dataset, obtaining a set of image feature vectors. This network model, through its regularized network architecture design and efficient feature representation capabilities, significantly optimizes the accuracy, efficiency, and generalization of the retrieval task, making it more suitable for large-scale, cross-scene image retrieval needs. Based on the aforementioned set of image feature vectors, an optimized clustering algorithm is used to perform clustering operations, obtaining a hierarchical structure. Given two integer parameters... and ,in This indicates the number of cluster centers, while This is used to limit the maximum number of feature vectors contained in each cluster. A common method for generating index trees using the K-means algorithm is to set... and Taking the same value would be a rigid approach. The value affects the clustering process of each branch, thus negatively impacting retrieval accuracy. Therefore, to improve the algorithm's adaptability in real-world scenarios and enhance retrieval accuracy, this invention proposes introducing a control parameter. This is designed to flexibly adjust the number of vectors in a cluster when building the index tree. Parameters The settings ensure that the number of nodes in each cluster is not excessive at a finer granular level, which helps to improve the clustering quality and make the number of nodes in each cluster relatively balanced.

[0037] The process of dividing the feature vector set into K different clusters using the K-means algorithm is as follows: First, randomly select... Each feature vector is used as the initial cluster center vector; then, the distance between each feature vector and each cluster center vector is calculated, and it is assigned to the nearest cluster. These two steps are performed iteratively until a preset number of iterations is reached or the cluster center vectors tend to stabilize.

[0038] After the initial partitioning described above, the number of vectors in each subspace may still be too large. At this point, to ensure that the number of vectors in each subspace does not exceed a set value... The K-means algorithm will be executed again on the corresponding subspaces containing too many vectors. This partitioning process will continue until all subspaces meet the requirements. Finally, all feature vectors that are leaf nodes are clustered under a root node, thus constructing a hierarchical structure based on feature clustering, where each leaf node corresponds to a feature vector of an image. This algorithm uses parameters... By limiting the maximum capacity of each cluster and automatically triggering recursive clustering when the number of samples within a cluster exceeds a threshold, the system ensures a relatively balanced number of samples across subspaces. This dynamic adjustment mechanism significantly improves the similarity of samples within clusters, laying a solid foundation for efficient subsequent retrieval. Simultaneously, by establishing a hierarchical tree-like index, the system avoids traversing all data during queries, instead quickly locating the target region along the optimal branch path. This reduces retrieval complexity from linear to logarithmic levels, minimizing unnecessary computational overhead while maintaining high recall.

[0039] For example, the system has 28 image feature vectors, set ,like Figure 4 As shown in (a), circles represent feature vectors, and squares represent cluster center vectors. After one K-means algorithm, all feature vectors are divided into 3 clusters, as follows: Figure 3 In (b), for features with a number greater than 100, 200, 300, 400, 500, 600, 7 ... The clustering is performed again using the K-means algorithm until the number of feature vectors in each cluster does not exceed a certain threshold. ,like Figure 3 As shown in (c) in the figure.

[0040] Finally, the image data to be retrieved is encrypted using the AES algorithm to obtain encrypted images, which are then uploaded to the cloud server. The AES encryption algorithm, through the high security and efficiency of symmetric cryptography, provides end-to-end privacy protection for the storage, transmission, and retrieval of image data, while also meeting the real-time and compatibility requirements of retrieval systems.

[0041] 2) The image owner uses an enhanced feature encryption algorithm to encrypt the hierarchical structure, resulting in a searchable index tree, achieving a balance between security and practicality. Firstly, for the... Zhang Image eigenvectors Encryption is performed using an enhanced feature-based encryption algorithm, based on the feature vector. Dimensions Generate key set ,in, It is a random permutation sequence used to shuffle the positions of the elements within the feature vector. It is a matrix and its invertible matrix. For eigenvectors... The feature vector is expanded to obtain the expanded feature vector. The expression is: ,in, It is an eigenvector The sum of the squares of all elements It is a random number; in this example, it is set to... =1.0. For the expanded feature vector... Using key sets Encryption is performed to obtain the encrypted feature vector. The expression is: .

[0042] The random permutation sequence aims to completely destroy the structure of the ciphertext. Even if an attacker can somehow deduce the meaning of a particular ciphertext component, they cannot determine which dimension that component corresponds to in the original feature vector. This significantly increases the difficulty of cryptanalysis, making it extremely difficult for attackers to deduce the encryption matrix from the ciphertext. Alternatively, understanding the semantic structure of feature vectors can effectively protect the patterns and distributions of the original features. Extending the feature vectors by incorporating random numbers ensures that each encryption result for the same image is different, preventing attackers from performing frequency analysis through repeated ciphertexts. More importantly, it achieves unlinkability of the ciphertexts, meaning the server cannot determine whether two different ciphertexts originate from the same original feature, thus effectively protecting the privacy of data association.

[0043] Based on the enhanced feature encryption algorithm described above, each node in the hierarchical structure is encrypted to obtain a searchable index tree. By uniformly applying the enhanced encryption algorithm to each node in the hierarchical structure, the entire index tree is transformed into a completely encrypted searchable structure. This means that when the cloud server performs a retrieval, every comparison and path selection from the root node to the leaf node is performed on encrypted data. It can only advance the search based on the encrypted comparison operations allowed by the encryption algorithm, and cannot know what the nodes it is comparing specifically represent, thus cutting off the chain of privacy leakage at the source.

[0044] like Figure 3 As shown in (d) above, this is the searchable index tree for this example. Figure 3 The four processes shown are schematic diagrams illustrating the process of generating the searchable index tree in this example.

[0045] 3) The query user extracts image features from the image data to be retrieved, obtaining the feature vector of the image to be retrieved. This feature vector is then encrypted using an enhanced feature encryption algorithm to obtain a searchable trapdoor, which is then uploaded to the cloud server. The RegNet convolutional neural network model is used to extract image features from the image data to be retrieved, obtaining the feature vector of the image to be retrieved. The feature vector generated by the above neural network model is not only compact but also effectively captures the key visual content of the image, thus achieving fast and accurate similarity matching in tasks such as image retrieval. The enhanced feature encryption algorithm is used to encrypt the feature vector of the image to be retrieved, obtaining a searchable trapdoor, as detailed below: Image feature vector to be retrieved The feature vector of the image to be retrieved is then expanded. The expression is: ,in, It is a random number. It is a positive integer and is different in each query. This applies to the expanded feature vector of the image to be retrieved. Using key sets Encryption is performed to obtain a searchable trapdoor. The expression is: ,in, yes The transpose of the matrix. The aforementioned expansion and encryption processing of the feature vector of the image to be retrieved is to construct a cryptographic trapdoor that can securely hide the query content while achieving accurate comparison in the ciphertext domain. The three independent random numbers introduced, especially the unique and positive integers for each query... This ensures that even when repeatedly querying the same image, the generated trapdoors will be completely different each time. This randomization effectively prevents cloud servers from inferring user search intent or identifying duplicate queries by analyzing trapdoor patterns, thus protecting user search pattern privacy. Simultaneously, the random factor introduced into the extended vector... The random number used in hierarchical encryption By working together, necessary noise barriers are reduced in subsequent ciphertext similarity calculations. After algebraic operations, the Euclidean distance between the original feature vectors can be restored, thus achieving the same ranking result as plaintext retrieval without revealing any plaintext information.

[0046] For example, setting For these three random numbers, it is only necessary to ensure The value should be a positive integer and different in each query. This is to ensure the validity of the Euclidean distance in the ciphertext field during subsequent retrievals. If the result is not a positive integer, Euclidean distance comparison will be impossible, causing similarity retrieval to fail.

[0047] 4) The cloud server uses Euclidean distance as the retrieval benchmark, calculates the searchable trapdoors and searchable index tree to obtain similar encrypted images in ciphertext state, and then sends them to the querying user. The querying user uses the decryption key sent by the image owner to decrypt the similar encrypted images in ciphertext state, obtaining the similar images in the image dataset. The search begins from the root node of the searchable index tree, and at each level, the inner product value between all node vectors at that level and the searchable trapdoors is calculated. This value, in its encrypted state, represents the Euclidean distance, expressed as:

[0048] in, It is a node vector of a searchable index tree. It is a searchable trapdoor. It is a feature vector in plaintext state and the feature vector of the image to be retrieved The squared Euclidean distance between the vectors is calculated. By employing algebraic transformations, the Euclidean distance is converted into a dot product operation of encrypted vectors, cleverly achieving efficient similarity calculation in the ciphertext domain. The calculation process does not require decrypting sensitive data, directly outputting a numerical result proportional to the plaintext distance while protecting query and index privacy, thus balancing retrieval accuracy and data security.

[0049] The distance values ​​are sorted, and the node with the smallest Euclidean distance (i.e., the most similar node) is selected as the path to proceed to the next level. This process is repeated until a leaf node is reached, indicating that the most similar encrypted image has been found. This sorting mechanism achieves efficient encrypted text retrieval by quantifying similarity. Minimizing the distance directly corresponds to the nearest encrypted image in the feature space, ensuring consistency between the retrieval results and the plaintext domain.

[0050] The set of similar encrypted images in the above-mentioned ciphertext state is sent to the querying user. The querying user uses the decryption key sent by the image owner to decrypt the set of similar encrypted images in the ciphertext state to obtain the required set of similar images.

[0051] For example, to further verify the effectiveness of this embodiment, the accuracy of image retrieval will be explained. The Corel-10k image dataset is used as the image data set of the image owners, and then 1000 images are randomly selected from it as the set of images to be retrieved owned by the query user. The Corel-10k image dataset contains 10,000 images, with 100 image categories, and each category contains 100 similar images. As an indicator of accuracy, among them, This indicates the number of similar encrypted images in the ciphertext state returned by the cloud server. Indicates the number of similar images. The closer the value is to 1, the higher the accuracy of the search.

[0052] Table 1. Comparison of image retrieval accuracy under different clustering parameters ; Table 1 shows the clustering parameters. When incrementing from 50 to 250, set the number of images returned. When the clustering parameter is 10, the accuracy of image retrieval is compared. Table 1 shows the results. Achieving the highest retrieval precision at this time indicates that the number of clusters... and node capacity It is correct and necessary to treat them as two independent control parameters. Setting them to the same value as in traditional methods would make it difficult to find the optimal balance. The existence of this peak confirms that the optimized clustering algorithm proposed in this invention can effectively find the configuration that best suits the index tree structure for a specific dataset by adjusting these two parameters, thus fundamentally guaranteeing the high practicality and deployability of this invention in real-world scenarios. Specific Implementation Example 2: like Figure 3 The diagram illustrates the execution flow of an image retrieval method based on polynomial searchability, with modifications made to the image encryption method. The detailed flow is as follows: 1) The image owner extracts image features from the image dataset to obtain a set of image feature vectors. Based on the similarity of image features, an optimized clustering algorithm is used to cluster the image feature vector set, resulting in a hierarchical structure of the image dataset. The image dataset is then encrypted using an image encryption algorithm based on compressed sensing and block DNA encoding to obtain an encrypted image. A convolutional neural network model is used to extract image features from the image dataset, obtaining another set of image feature vectors. Using this set of image feature vectors, an optimized clustering algorithm is used to obtain a hierarchical structure based on the similarity of image features. The image data to be retrieved is then encrypted using an image encryption algorithm based on compressed sensing and block DNA encoding to obtain an encrypted image, which is then uploaded to a cloud server. The image encryption method based on compressed sensing and block DNA encoding first obtains the chaotic sequence used in the encryption process and constructs a measurement matrix through an iterative composite chaotic system. Compressed sensing is used to compress the color image to obtain three measurement value matrices, which are then quantized into integer matrices and scrambled using a confusion method based on the Josephus problem. Subsequently, these scrambled measurement value matrices are divided into sub-blocks. The sub-blocks undergo dynamic DNA encoding, DNA diffusion, and DNA decoding, and are then recombined to obtain the encrypted image. This method compresses the plaintext image to obtain a small-sized encrypted image, while ensuring that the encrypted image can resist attacks such as noise attacks, blocking attacks, statistical attacks, chosen-plaintext attacks, and known-plaintext attacks during transmission.

[0054] 2) The image owner uses an enhanced feature encryption algorithm to encrypt the above hierarchical structure, resulting in a searchable index tree.

[0055] 3) The query user extracts image features from the image data to be retrieved, obtains the feature vector of the image to be retrieved, and encrypts the feature vector of the image to be retrieved according to the enhanced feature encryption algorithm to obtain the searchable trapdoor, and uploads the searchable trapdoor to the cloud server.

[0056] 4) The cloud server uses Euclidean distance as the retrieval benchmark to calculate the searchable trapdoors and searchable index trees to obtain similar encrypted images in ciphertext state, which are then sent to the querying user. The querying user uses the decryption key sent by the image owner to decrypt the similar encrypted images in ciphertext state, obtaining the similar images from the image dataset. Specific Implementation Example 3: Further modifications were made to the image feature extraction method. The detailed process is as follows: 1) The image owner extracts image features from the image dataset, obtaining a set of image feature vectors. Based on the similarity of image features, an optimized clustering algorithm is used to cluster the image feature vector set, resulting in a hierarchical structure of the image dataset. The image dataset is then encrypted using the AES algorithm to obtain an encrypted image. First, the image owner uses the DenseNet convolutional neural network model to extract image features from the image dataset, obtaining a set of image feature vectors. This network model deeply integrates shallow detail textures, mid-level object parts, and deep semantic concepts, ensuring that the final extracted global image features simultaneously contain detailed local information and high-level semantic information. This enables more accurate similarity matching when faced with deformation, occlusion, or complex backgrounds. Using the aforementioned set of image feature vectors, an optimized clustering algorithm is used to obtain a hierarchical structure based on the similarity of image features. The image data to be retrieved is then encrypted using the AES algorithm to obtain an encrypted image, which is then uploaded to a cloud server.

[0058] 2) The image owner uses an enhanced feature encryption algorithm to encrypt the hierarchical structure, resulting in a searchable index tree.

[0059] 3) The query user extracts image features from the image data to be retrieved, obtaining the feature vector of the image to be retrieved. This feature vector is then encrypted using an enhanced feature encryption algorithm to obtain a searchable trapdoor, which is then uploaded to the cloud server. The DenseNet convolutional neural network model is used to extract image features from the image data to be retrieved, obtaining the feature vector of the image data to be retrieved; the feature vector of the image to be retrieved... The feature vector of the image to be retrieved is then expanded. The expression is: ,in, It is a random number. It is a positive integer and is different in each query. This applies to the expanded feature vector of the image to be retrieved. Using key sets Encryption is performed to obtain a searchable trapdoor. The expression is: ,in, yes The transpose of the matrix is ​​then used. The searchable trapdoors are then uploaded to the cloud server.

[0060] 4) The cloud server uses Euclidean distance as the retrieval benchmark to calculate the searchable trapdoors and searchable index trees to obtain similar encrypted images in ciphertext state, which are then sent to the querying user. The querying user uses the decryption key sent by the image owner to decrypt the similar encrypted images in ciphertext state, obtaining the similar images from the image dataset.

[0061] For example, to further verify the effectiveness of this embodiment, the accuracy of image retrieval will be explained. The Corel-10k image dataset is used as the image data set of the image owners, and then 1000 images are randomly selected from it as the set of images to be retrieved owned by the query user. The Corel-10k image dataset contains 10,000 images, with 100 image categories, and each category contains 100 similar images. As an indicator of accuracy, among them, This indicates the number of similar encrypted images in the ciphertext state returned by the cloud server. Indicates the number of similar images. The closer the value is to 1, the higher the accuracy of the search.

[0062] Table 2 Comparison of image retrieval accuracy under different image feature extraction methods ; In Table 2, when the number of returned images is... As the value increases from 10 to 50, the accuracy of image retrieval using two image feature extraction methods is compared. DenseNet represents feature extraction from image data using the DenseNet neural network model, while CLD represents extracting the color layout descriptor of the image as its feature vector. Table 2 shows that the retrieval accuracy achieved using the DenseNet image feature extraction method is higher than that achieved using the CLD image feature extraction method. This indicates that, by leveraging a convolutional neural network pre-trained on a large-scale image set and fine-tuning it in this invention, it can deeply mine the deep semantic features of images, more accurately reflecting the true content of the images. Compared to traditional local and global feature retrieval methods, this invention significantly improves retrieval accuracy and better meets practical application needs.

[0063] Reference Figure 2 An image retrieval system based on improved feature clustering includes: The first module 201 is used to perform clustering and encryption operations on the image data set to obtain the hierarchical structure of the image data set and the encrypted image. The second module 202 is used to encrypt the hierarchical structure of the image data set through an enhanced feature encryption algorithm to obtain a searchable index tree; The third module 203 is used to sequentially extract and encrypt image features from the image data to be retrieved, thereby obtaining a searchable trapdoor; The fourth module 204 is used to use Euclidean distance as a retrieval benchmark to calculate the searchable trapdoors and searchable index trees, obtain similar images from the image dataset, and decrypt encrypted images to obtain similar images from the image dataset.

[0064] The content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0065] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. An image retrieval method based on improved feature clustering, characterized in that, Includes the following steps: Clustering and encryption operations are performed on the image dataset to obtain the hierarchical structure of the image dataset and the encrypted image. The hierarchical structure of the image data set is encrypted using an enhanced feature encryption algorithm to obtain a searchable index tree; Image features are extracted and encrypted sequentially from the image data to be retrieved to obtain searchable trapdoors; Using Euclidean distance as the retrieval benchmark, the searchable trapdoors and searchable index trees are calculated to obtain similar images from the image dataset and to decrypt encrypted images, thus obtaining similar images from the image dataset.

2. The image retrieval method based on improved feature clustering according to claim 1, characterized in that, The step of performing clustering and encryption operations on the image data set to obtain the hierarchical structure of the image data set and the encrypted image specifically includes: Image features are extracted from the image dataset using the Regnet convolutional neural network model, resulting in a set of image feature vectors. Based on the similarity of image features, an optimized clustering algorithm is used to cluster the image feature vector set to obtain the hierarchical structure of the image data set. The image data set is encrypted using the AES algorithm to obtain an encrypted image.

3. The image retrieval method based on improved feature clustering according to claim 2, characterized in that, The step of clustering the image feature vector set based on the similarity of image features using an optimized clustering algorithm to obtain the hierarchical structure of the image data set specifically includes: The number of cluster centers is set and a control parameter is introduced, which is used to limit the maximum number of feature vectors contained in each cluster; Based on the number of cluster centers, select several vectors as initial cluster center vectors, and calculate the distance between the image feature vector and each initial cluster center vector; Based on the calculated distance value, the image feature vector is divided to the nearest initial cluster center vector until the entire image feature vector set is traversed to obtain the initially divided image feature vector set, which includes several subspaces; The image feature vector set after initial partitioning is filtered according to the control parameters. Subspaces with a feature vector number greater than a preset threshold are selected for re-partitioning until the feature vector number of all subspaces meets the control parameters, and the hierarchical structure of the image data set is output.

4. The image retrieval method based on improved feature clustering according to claim 3, characterized in that, The step of encrypting the hierarchical structure of the image data set using an enhanced feature encryption algorithm to obtain a searchable index tree specifically includes: Generate a key set based on the dimension of the feature vectors in the hierarchical structure of the image dataset; The feature vectors in the hierarchical structure of the image dataset are expanded to obtain expanded feature vectors. The expanded feature vector is encrypted using a key set to obtain the encrypted feature vector. Repeat the expansion and encryption steps until all nodes in the hierarchical structure of the image data set are traversed, and a searchable index tree is constructed.

5. The image retrieval method based on improved feature clustering according to claim 4, characterized in that, The step of sequentially extracting and encrypting image features from the image data to be retrieved to obtain a searchable trapdoor specifically includes: The Regnet convolutional neural network model is used to extract image features from the image data to be retrieved, resulting in the feature vector of the image to be retrieved. The feature vector of the image to be retrieved is expanded to obtain the expanded feature vector of the image to be retrieved. By encrypting the expanded feature vector of the image to be retrieved using a key set, a searchable trapdoor can be obtained.

6. The image retrieval method based on improved feature clustering according to claim 5, characterized in that, The step of using Euclidean distance as a retrieval benchmark to calculate searchable trapdoors and searchable index trees, obtaining similar images from the image dataset, and decrypting encrypted images to obtain similar images from the image dataset specifically includes: Search the root node of the searchable index tree. At each level, calculate the inner product between all node vectors at that level and the searchable trapdoors to obtain the Euclidean distance value. Sort the Euclidean distance values, select the node with the smallest Euclidean distance value as the path, proceed to the next level, until the leaf node of the searchable index tree is reached, and output the most similar encrypted image. Based on the most similar encrypted image, the encrypted image is decrypted using the AES algorithm to obtain the similar image in the image dataset.

7. The image retrieval method based on improved feature clustering according to claim 6, characterized in that, The specific expression for calculating the Euclidean distance value is as follows: ； In the above formula, This represents the inner product value between all node vectors and searchable trapdoors. Represents the Euclidean distance value. This represents a vector of a node in a searchable index tree. This indicates a searchable trapdoor.

8. An image retrieval system based on improved feature clustering, characterized in that, Includes the following modules: The first module is used to perform clustering and encryption operations on the image data set to obtain the hierarchical structure of the image data set and the encrypted image. The second module is used to encrypt the hierarchical structure of the image data set through an enhanced feature encryption algorithm to obtain a searchable index tree. The third module is used to extract and encrypt the image data to be retrieved in sequence to obtain searchable trapdoors; The fourth module is used to use Euclidean distance as a retrieval benchmark to calculate the searchable trapdoors and searchable index trees, obtain similar images from the image dataset, and decrypt encrypted images to obtain similar images from the image dataset.

Images