Image retrieval method and system based on multi-dimensional features

Abstract

The invention discloses an image retrieval method and system based on multi-dimensional features, and relates to the technical field of image retrieval. The method comprises the following steps: processing a to-be-retrieved image and an image in a database to generate a multi-dimensional feature vector set, and constructing a to-be-analyzed feature sequence; calculating feature change values, feature level values and neighborhood feature differences of the feature vectors, determining feature difference coefficients, presetting suspected parameters T1, and screening out suspected matching feature vectors with the feature difference coefficients greater than or equal to T1; calculating matching selection factors and screening matching feature vectors; carrying out similarity evaluation and division evaluation value calculation on the feature data blocks; determining the optimal division length; and performing hash matching and outputting a retrieval result. According to the method, through the multi-dimensional feature extraction and dynamic division technology, global and local feature analysis is combined, and the image retrieval precision and efficiency are improved.

Description

Technical Field

[0001] The present invention belongs to the field of image retrieval technology, and in particular relates to an image retrieval method and system based on multi-dimensional features. Background Art

[0002] With the rapid development of the Internet, image data has exploded. How to efficiently and accurately retrieve images of interest to users from massive images has become an urgent problem to be solved. Traditional text-based image retrieval methods, such as retrieval by keywords, tags, etc., have the problem of semantic gap, that is, there is often a mismatch between the visual content of the image and the text description, resulting in poor retrieval results. Therefore, image content-based image retrieval methods came into being. They perform retrieval by analyzing the image content features and subject object features (such as color, shape, texture, structure, etc.), which can more directly reflect the visual content of the image and improve the accuracy of retrieval.

[0003] Although image content-based image retrieval methods have made some progress, there are still some limitations: Single feature extraction: Many existing image retrieval methods only use a single feature extraction algorithm, such as color features, shape features, or texture features. However, a single feature is often difficult to fully describe the content of the image, especially when facing complex scenes or images with multiple features, the retrieval effect will be limited.

[0004] Inaccurate feature matching: In the feature matching stage, existing methods may rely too much on simple similarity metrics, such as Euclidean distance, cosine similarity, etc. These methods may not be accurate enough when processing images with complex features. In addition, insufficient consideration of the spatial distribution and semantic relevance of features may also lead to inaccurate matching results.

[0005] Low retrieval efficiency: When faced with massive image data, the retrieval efficiency of existing methods is often low, making it difficult to meet the needs of real-time retrieval.

[0006] This is mainly due to the high computational complexity of steps such as feature extraction and feature matching, which leads to slow retrieval speed.

[0007] In response to the above limitations, this invention patent proposes an image retrieval method based on multi-dimensional features, aiming to achieve more efficient and accurate image retrieval through the steps of multi-dimensional feature extraction and feature vector construction, feature difference coefficient calculation and suspected matching feature vector screening, matching selection factor calculation and matching feature vector screening, feature data block similarity evaluation and partition evaluation value calculation, optimal partition length determination, hash matching and retrieval result output. Summary of the invention

[0008] In order to overcome the shortcomings and deficiencies of the above-mentioned prior art, the first purpose of the present invention is to provide an image retrieval method based on multi-dimensional features; the second purpose of the present invention is to provide an image retrieval system based on multi-dimensional features.

[0009] The first object of the present invention adopts the following technical solution: The image retrieval method based on multi-dimensional features has the following process: Step 1: Process the images to be retrieved and the images in the database to generate a multi-dimensional feature vector set and construct a feature sequence to be analyzed; Step 2: Calculate feature difference coefficient and screen suspected matching feature vectors: Calculate feature change value, feature level value, neighborhood feature difference of feature vectors, determine feature difference coefficient, preset suspected parameter T1, and screen suspected matching feature vectors whose feature difference coefficient is greater than or equal to T1; Step 3: Calculate the matching selection factor and screen the matching feature vector: Calculate the neighborhood direction similarity and feature prominence eigenvalue of the suspected matching feature vector, determine the matching selection factor, preset the selection parameter T2, and screen out the matching feature vector whose matching selection factor is greater than or equal to T2; Step 4: Feature data block similarity evaluation and partition evaluation value calculation: Improve the feature data block similarity measurement, calculate the global similar feature value, screen the difference data block, and calculate the partition evaluation value; Step 5: Determine the optimal partition length: Calculate the partition effect characterization value and the partition effect slope value at each length, determine the characteristic evaluation coefficient and the degree of preference, and take the minimum length corresponding to the degree of preference greater than the preset threshold as the optimal length; Step 6. Hash matching and retrieval result output: Optimize the hash function, select the seed data block for matching and calculate the matching score, determine the retrieval result based on the matching score and filter the output.

[0010] Preferably, the feature change value is the absolute value of the difference between the feature values ​​of the feature vector in the image to be retrieved and each image in the image database; the feature level value is the difference between the maximum eigenvalue of the image where the feature vector is located and the eigenvalue of the feature vector; the neighborhood feature difference is the average of the feature difference values ​​between all local feature vectors within the 8-neighborhood range of the feature vector and the feature vector.

[0011] Preferably, the characteristic difference coefficient is determined by normalizing the ratio of the characteristic change value to the characteristic level value of the characteristic vector.

[0012] Preferably, the method for calculating the neighborhood direction similarity is: taking all local feature vectors within the 8-neighborhood range of the suspected matching feature vector as reference feature vectors, and calculating the average of the cosine values ​​of the angles between the maximum direction of the suspected matching feature vector and the maximum direction of each reference feature vector.

[0013] Preferably, the feature highlighting feature value is the ratio of the feature difference coefficient of the suspected matching feature vector to the minimum value of the feature difference coefficients of all suspected matching feature vectors in the image database.

[0014] Preferably, the matching selection factor is determined by normalizing the product of the neighborhood direction similarity of the suspected matching feature vector and the feature salient feature value.

[0015] Preferably, the method for improving the feature data block similarity measurement is as follows: in addition to using the DTW distance to measure the similarity of the position correspondence of the feature data blocks, a feature vector similarity measurement based on cosine similarity is introduced, and the spatial context information of the feature data blocks is considered, and the similarity weighted and optimized result of the surrounding neighborhood feature data blocks is calculated.

[0016] Preferably, the method for calculating the global similarity feature value is: combining DTW distance, cosine similarity and spatial context information to determine the comprehensive similarity feature factor between the target data block and the reference data block, and after normalization, determining it based on the number of reference data blocks greater than a preset similarity threshold combined with a weighted strategy.

[0017] Preferably, the hash function optimization adopts a multi-hash function combination strategy based on local sensitive hashing, designs corresponding hash functions for different types of feature data blocks, and adjusts parameters so that similar feature data blocks are mapped to the same or similar hash buckets.

[0018] The second object of the present invention adopts the following technical solution: An image retrieval system based on multi-dimensional features is used to implement an image retrieval method based on multi-dimensional features. The system includes a feature extraction construction module, a feature difference coefficient calculation and screening module, a matching selection factor calculation and screening module, a feature data block similarity evaluation and division module, an optimal division length determination module, and a hash matching retrieval result output module; Feature extraction and construction module: used to extract multi-dimensional features from the images to be retrieved and the images in the database, and to construct feature vectors and feature sequences; Feature difference coefficient calculation and screening module: used to calculate the difference coefficient of feature vectors and screen out suspected matching feature vectors; Matching selection factor calculation and screening module: used to calculate the matching selection factor of the suspected matching feature vector and screen out the matching feature vector; Feature data block similarity evaluation and partitioning module: used to evaluate the similarity of feature data blocks and calculate the partitioning evaluation value to determine the optimal partitioning length; Optimal partition length determination module: used to determine the optimal partition length according to the partition evaluation value; Hash matching retrieval result output module: used to use hash functions for fast matching and output retrieval results.

[0019] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. The present invention can comprehensively measure the degree of difference between features in different images and the uniqueness in local areas by integrating multiple feature measurement indicators. The calculation method of the feature difference coefficient and the matching selection factor combines the global and local information of the features, so that the screened feature vectors are more discriminative and representative, and can more accurately find features that match the image to be retrieved. The improved similarity measurement method comprehensively considers the position correspondence, value distribution, direction information and spatial context information of the features, further improving the accuracy of the similarity evaluation of the feature data blocks, thereby improving the overall image retrieval matching accuracy; 2. The present invention can adaptively adjust the window size and sliding step size according to the characteristics of the feature through the adaptive sliding window technology, better capture the local and global characteristics of the feature, and make the division of the feature data block more reasonable. The multi-hash function combination strategy designs corresponding hash functions for different types of feature data blocks, and maps similar feature data blocks to the same or similar hash buckets by adjusting parameters, which greatly reduces the amount of calculation in the matching process and improves the retrieval efficiency. In addition, through dynamic division and hash matching, it is also possible to quickly screen out image feature templates that match the image to be retrieved while ensuring the accuracy of the retrieval. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0021] Figure 1 A flowchart of the image retrieval method based on multi-dimensional features of the present invention is shown; Figure 2 The module diagram of the image retrieval system based on multi-dimensional features of the present invention is shown; Figure 3 A flow chart of the characteristic difference coefficient calculation and suspected matching characteristic vector screening of the present invention is shown. DETAILED DESCRIPTION

[0022] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0023] In addition, the described features, structures or characteristics may be combined in one or more example embodiments in any suitable manner. In the following description, many specific details are provided to provide a full understanding of the example embodiments of the present disclosure. However, those skilled in the art will appreciate that the technical solutions of the present disclosure may be practiced while omitting one or more of the specific details, or other methods, components, steps, etc. may be adopted. In other cases, well-known structures, methods, implementations or operations are not shown or described in detail to avoid obscuring various aspects of the present disclosure.

[0024] Embodiment 1: See also Figure 1 As shown, the image retrieval method based on multi-dimensional features in this embodiment has the following process: Step 1: Multi-dimensional feature extraction and feature vector construction.

[0025] For the image to be retrieved, a variety of complementary feature extraction algorithms are used to extract local features using SIFT to generate feature vectors ; Use HOG to extract shape features and generate feature vectors ; Use the convolutional neural network (CNN) based on deep learning to extract high-level semantic features and generate feature vectors , forming a multi-dimensional feature vector set, combining these feature vectors in a specific order to construct the feature sequence to be analyzed .

[0026] For the image feature templates in the database, the same feature extraction algorithm is used to generate a multi-dimensional original feature sequence .

[0027] Each set of features is processed as follows: According to the spatial distribution and semantic relevance of the features, each set of features is dynamically divided using adaptive sliding window technology to obtain feature data blocks of each original feature sequence of different lengths. The window size and sliding step size are adaptively adjusted according to the characteristics of the features to better capture the local and global characteristics of the features.

[0028] Step 2: Calculate feature difference coefficient and screen suspected matching feature vectors.

[0029] See also Figure 3As shown in the figure, the process of calculating the feature difference coefficient and screening the suspected matching feature vector is as follows: S21, feature change value calculation: the absolute value of the difference between the feature values ​​of the feature vector in the image to be retrieved and each image in the image database is recorded as the feature change value of the feature vector: ; It should be noted that the feature change value reflects the degree of difference between the feature vectors in the image to be retrieved and the images in the image database.

[0030] S22, characteristic level value calculation: the difference between the maximum eigenvalue of the image where the characteristic vector is located and the eigenvalue of the characteristic vector is recorded as the characteristic level value of the characteristic vector: ; It should be noted that the feature level value reflects the relative feature level of the feature vector in the image where it is located.

[0031] S23, neighborhood feature difference calculation: The feature vector All local feature vectors within the 8-neighborhood range are recorded as the contrast feature vectors of the feature vector. For any contrast feature vector, the absolute value of the difference between the feature vector and the feature value of the contrast feature vector is recorded as the feature difference value of the contrast feature vector. The mean of the feature difference values ​​of all contrast feature vectors is taken as the neighborhood feature difference of the feature vector: ; in, is the neighborhood feature vector, is the number of neighborhood eigenvectors.

[0032] It should be noted that the ratio of the feature change value to the feature level value (the first ratio) can measure the relative size of the change of the feature between different images relative to the feature level in its own image. Mathematically, when the feature change value is large and the feature level value is small, the first ratio will be large, indicating that the change of the feature between different images is more significant relative to the feature level of its own image, and has greater potential for differentiation.

[0033] The neighborhood feature difference measures the difference between the feature vector and the local feature vector within its 8 neighborhoods, reflecting the uniqueness of the feature in the local area. If the neighborhood feature difference of a feature vector is large, it means that it is significantly different from the surrounding features in the local area and has a stronger local distinction ability.

[0034] Multiplying the first ratio by the neighborhood feature difference is a weighted combination method that can comprehensively consider the global (between different images) and local (within the neighborhood) changes of the feature. In image retrieval, we need to pay attention to both the global and local information of the feature. If a feature has a large global change and is unique locally, it will be more capable of distinguishing different images. Through this multiplication method, the degree of difference of the feature can be more comprehensively measured, so as to more accurately screen out the features that may match the image to be retrieved. Finally, the multiplication result is normalized to unify the feature difference coefficient within a reasonable range to facilitate subsequent comparison and screening.

[0035] S24, feature difference coefficient determination: the ratio of the feature change value of the feature vector to the feature level value of the feature vector is recorded as the first ratio: ; Will Normalization is performed to obtain the characteristic difference coefficient of the eigenvector . S25, suspected matching feature vector screening: preset a suspected parameter T1, for any feature vector in the image database, if its feature difference coefficient If it is greater than or equal to the suspected parameter T1, the feature vector is marked as a suspected matching feature vector. It should be noted that the suspected parameter is a preset threshold used to filter suspected matching feature vectors. Specifically, it is a standard used to determine whether a feature vector is likely to match the image to be retrieved.

[0036] Step 3: Calculate matching selection factors and screen matching feature vectors.

[0037] S31. Neighborhood directional similarity calculation: All local feature vectors within the 8-neighborhood range of the suspected matching feature vector are recorded as the reference feature vectors of the suspected matching feature vector. For any reference feature vector, the cosine value of the angle between the maximum direction of the suspected matching feature vector and the maximum direction of the reference feature vector is recorded as the directional similarity value of the reference feature vector; the average of the directional similarity values ​​of all reference feature vectors is taken as the neighborhood directional similarity of the suspected matching feature vector: ; in, and They are and The maximum direction.

[0038] S32, feature salient eigenvalue calculation: the ratio of the feature difference coefficient of the suspected matching feature vector to the minimum value of the feature difference coefficients of all suspected matching feature vectors in the image database is used as the feature salient eigenvalue of the suspected matching feature vector: .

[0039] S33, determining the matching selection factor: normalizing the product of the neighborhood direction similarity of the suspected matching feature vector and the feature prominence feature value, and using the value as the matching selection factor of the suspected matching feature vector: ; right Perform normalization to obtain the final matching selection factor .

[0040] S34, matching feature vector screening: preset a selection parameter T2, for any suspected matching feature vector in the image database, if its matching selection factor If it is greater than or equal to the selection parameter T2, the suspected matching feature vector is marked as a matching feature vector.

[0041] Step 4: Evaluate the similarity of feature data blocks and calculate the partition evaluation value.

[0042] S41. Improved similarity metric: When calculating the similarity of feature data blocks, in addition to using the DTW distance to measure the similarity of the position correspondence of feature data blocks, a feature vector similarity metric based on cosine similarity is introduced to comprehensively consider the numerical distribution and direction information of the features. At the same time, the spatial context information of the feature data block is considered, and the similarity weighted sum of the feature data blocks in its surrounding neighborhood is calculated to further optimize the similarity metric result: ; in, , , is the weight coefficient; is the spatial context similarity.

[0043] S42. Calculation of global similarity feature value: Mark a feature data block in any original feature sequence as a target data block, and mark the feature data blocks in other original feature sequences that correspond to the target data block in spatial structure and semantics as reference data blocks. Determine the comprehensive similarity feature factor between the target data block and each reference data block by combining DTW distance, cosine similarity, and spatial context information. After normalizing the comprehensive similarity feature factor, determine the global similarity feature value of the target data block based on the number of reference data blocks greater than the preset similarity threshold and the weighting strategy (giving different weights to reference data blocks corresponding to different types of features to reflect the importance of different feature types): ; in, is the weight; is the number of reference data blocks.

[0044] S43, difference data block screening: not only based on global similarity feature values The difference data blocks are determined to be smaller than the preset difference threshold T3, and feature type analysis and semantic information judgment are also introduced. For example, for some feature data blocks with specific semantic meanings but low similarity, even if their global similarity feature values ​​do not meet the condition of being smaller than the difference threshold, they are included in the consideration range of the difference data blocks. At the same time, a machine learning classifier (such as a support vector machine SVM) is used to classify the feature data blocks. According to the classification results, if the feature data block belongs to a specific category and has low similarity, it is included in the consideration range of the difference data block.

[0045] S44, calculation of partition evaluation value: In any original feature sequence of a certain length, for each difference data block, based on the negative correlation coefficient of the maximum value of the comprehensive similarity feature factor between it and the corresponding reference data block, combined with the semantic importance weight of the feature, determine the partition feature factor of each difference data block. The partition feature factors of all difference data blocks in the original feature sequence are integrated, and the weights of different feature types in the overall feature sequence are considered to determine the partition evaluation value of the original feature sequence of the length: ; in, is the partition characteristic factor of the difference data block; is the weight.

[0046] Step 5: Determine the optimal partition length.

[0047] S51. Calculation of partition effect representation value: Calculate the weighted mean of the partition evaluation values ​​of all original feature sequences at each length (different weights are assigned according to the importance of different feature types in image retrieval) to obtain the partition effect representation value of each length: .

[0048] S52. Calculation of the slope value of the partitioning effect: in ascending order of length, based on the rate of change of the partitioning effect characterization value between each length and the adjacent previous length, combined with the change trend of different feature types (by analyzing the change of the partitioning evaluation values ​​corresponding to different feature types), determine the slope value of the partitioning effect of each length: .

[0049] S53, calculation of feature evaluation coefficient and preference degree: For any length, based on the negative correlation coefficient of the difference in the slope value of the division effect between the length and the adjacent previous length, combined with the stability index of different feature types (determined by analyzing the fluctuation of the division evaluation value of different feature types at different lengths), determine the feature evaluation coefficient of the length. According to the feature evaluation coefficient and the characterization value of the division effect of the length, combined with the weights of different feature types, calculate the preference degree of the length. The minimum length corresponding to the preference degree greater than the preset preference threshold is taken as the optimal length: ; according to And the partitioning effect characterization value R, calculate the degree of preference.

[0050] Step 6: Hash matching and retrieval result output.

[0051] S61. Hash function optimization: Adopt a multi-hash function combination strategy based on locality sensitive hashing (LSH). Design corresponding hash functions for different types of feature data blocks (such as local features, shape features, and semantic features): ; When constructing the hash table, the similarity range and distribution characteristics of the feature data blocks are fully considered, and by adjusting the parameters of the hash function, similar feature data blocks are mapped to the same or similar hash buckets as much as possible.

[0052] S62, seed data block selection and matching: For any image feature template, under the optimal length, the global similarity feature value, the semantic importance of the feature, and the spatial position distribution of the feature data block in the image are comprehensively considered, and a representative difference data block is selected as the seed data block. Based on the seed data block and each data block in the feature sequence division result to be analyzed, a multi-hash function match is performed and the matching score is calculated. The matching score not only considers the consistency of the hash bucket, but also combines the similarity measurement results between the feature data blocks for weighted calculation. The matching score calculation formula is as follows: ; S63, sorting by matching score and outputting search results: Based on the matching score of each image feature template, obtain the feature matching degree of each image feature template in the feature sequence to be analyzed (by normalizing the matching score and calculating it in combination with the weights of different feature types). Arrange in descending order according to the feature matching degree to determine the search results of the image feature template. At the same time, according to the number and matching degree of the search results, set a threshold for screening, and only output image feature templates with a high matching degree and meeting a certain number of requirements to improve the quality of the search results.

[0053] The beneficial effects of this embodiment include: improving retrieval accuracy through multi-dimensional features and comprehensive evaluation strategies; enhancing feature discrimination by utilizing feature difference coefficients and local characteristics; optimizing the partition length to ensure the best matching effect; and accelerating the retrieval process through hash matching to improve overall efficiency and user experience.

[0054] Embodiment 2: See also Figure 2 As shown, the image retrieval system based on multi-dimensional features of this embodiment includes a feature extraction construction module, a feature difference coefficient calculation and screening module, a matching selection factor calculation and screening module, a feature data block similarity evaluation and division module, an optimal division length determination module and a hash matching retrieval result output module.

[0055] Feature extraction construction module: used to extract multi-dimensional features from the images to be retrieved and the images in the database, and to construct feature vectors and feature sequences.

[0056] Feature difference coefficient calculation and screening module: used to calculate the difference coefficient of feature vectors and screen out suspected matching feature vectors.

[0057] Matching selection factor calculation and screening module: used to calculate the matching selection factor of the suspected matching feature vector and screen out the matching feature vector.

[0058] Feature data block similarity evaluation partitioning module: used to evaluate the similarity of feature data blocks and calculate the partitioning evaluation value to determine the optimal partitioning length.

[0059] Optimal partition length determination module: used to determine the optimal partition length according to the partition evaluation value.

[0060] Hash matching retrieval result output module: used to use hash functions for fast matching and output retrieval results.

[0061] The beneficial effects of this embodiment are: improving the accuracy and efficiency of image retrieval, ensuring matching accuracy through multi-dimensional feature extraction and screening; using hash functions to achieve fast matching and shorten retrieval time; the system is modularly designed to facilitate maintenance and expansion to meet different retrieval needs.

[0062] The weights of the present invention are used to measure the degree of influence of different factors or variables on a certain result or decision. The definition of weight refers to the value assigned to each factor when comparing and evaluating multiple factors to reflect its importance or priority. These weights can be determined according to specific circumstances and needs, and are usually jointly formulated and confirmed by professionals or relevant stakeholders. By setting weights reasonably, it can help programs or systems make decisions or predictions more accurately.

[0063] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

[0064] The preferred embodiments of the present invention disclosed above are only used to help explain the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to only specific implementation methods. Obviously, many modifications and changes can be made according to the content of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can understand and use the present invention well. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. An image retrieval method based on multi-dimensional features, characterized in that: The method flow is as follows: Step 1: Process the images to be retrieved and the images in the database to generate a multi-dimensional feature vector set and construct a feature sequence to be analyzed; Step 2: Calculate feature difference coefficient and screen suspected matching feature vectors: Calculate feature change value, feature level value, neighborhood feature difference of feature vectors, determine feature difference coefficient, preset suspected parameter T1, and screen suspected matching feature vectors whose feature difference coefficient is greater than or equal to T1; Step 3: Calculate the matching selection factor and screen the matching feature vector: Calculate the neighborhood direction similarity and feature prominence eigenvalue of the suspected matching feature vector, determine the matching selection factor, preset the selection parameter T2, and screen out the matching feature vector whose matching selection factor is greater than or equal to T2; Step 4: Feature data block similarity evaluation and partition evaluation value calculation: Improve the feature data block similarity measurement, calculate the global similarity feature value, screen the difference data block, and calculate the partition evaluation value; Step 5: Determine the optimal partition length: Calculate the partition effect characterization value and the partition effect slope value at each length, determine the characteristic evaluation coefficient and the degree of preference, and take the minimum length corresponding to the degree of preference greater than the preset threshold as the optimal length; Step 6. Hash matching and retrieval result output: Optimize the hash function, select the seed data block for matching and calculate the matching score, determine the retrieval result based on the matching score and filter the output.

2. The image retrieval method based on multi-dimensional features according to claim 1, characterized in that: The feature change value is the absolute value of the difference between the feature values ​​of the feature vector in the image to be retrieved and each image in the image database; the feature level value is the difference between the maximum eigenvalue of the image where the feature vector is located and the eigenvalue of the feature vector; the neighborhood feature difference is the average of the feature difference values ​​between all local feature vectors within the 8-neighborhood range of the feature vector and the feature vector.

3. The image retrieval method based on multi-dimensional features according to claim 1, characterized in that: The characteristic difference coefficient is determined by normalizing the ratio of the characteristic change value to the characteristic level value of the characteristic vector.

4. The image retrieval method based on multi-dimensional features according to claim 1, characterized in that: The method for calculating the neighborhood direction similarity is: taking all local feature vectors within the 8-neighborhood range of the suspected matching feature vector as reference feature vectors, and calculating the average of the cosine values ​​of the angles between the maximum direction of the suspected matching feature vector and the maximum direction of each reference feature vector.

5. The image retrieval method based on multi-dimensional features according to claim 1, characterized in that: The feature highlighting feature value is the ratio of the feature difference coefficient of the suspected matching feature vector to the minimum value of the feature difference coefficients of all suspected matching feature vectors in the image database.

6. The image retrieval method based on multi-dimensional features according to claim 1, characterized in that: The matching selection factor is determined by normalizing the product of the neighborhood direction similarity of the suspected matching feature vector and the feature salient feature value.

7. The image retrieval method based on multi-dimensional features according to claim 1, characterized in that: The improved method for feature data block similarity measurement is as follows: in addition to using DTW distance to measure the similarity of the position correspondence relationship of feature data blocks, a feature vector similarity measurement based on cosine similarity is introduced, and the spatial context information of the feature data block is considered, and the similarity weighted sum optimization result of the surrounding neighborhood feature data blocks is calculated.

8. The image retrieval method based on multi-dimensional features according to claim 1, characterized in that: The global similarity feature value is calculated by combining DTW distance, cosine similarity and spatial context information to determine the comprehensive similarity feature factor between the target data block and the reference data block, and after normalization, it is determined based on the number of reference data blocks greater than a preset similarity threshold and combined with a weighted strategy.

9. The image retrieval method based on multi-dimensional features according to claim 1, characterized in that: The hash function optimization adopts a multi-hash function combination strategy based on local sensitive hashing, designs corresponding hash functions for different types of feature data blocks, and adjusts parameters so that similar feature data blocks are mapped to the same or similar hash buckets.

10. An image retrieval system based on multi-dimensional features, used to implement the image retrieval method based on multi-dimensional features as claimed in claim 1, characterized in that: The system includes a feature extraction construction module, a feature difference coefficient calculation and screening module, a matching selection factor calculation and screening module, a feature data block similarity evaluation and division module, an optimal division length determination module and a hash matching retrieval result output module; Feature extraction and construction module: used to extract multi-dimensional features from the images to be retrieved and the images in the database, and to construct feature vectors and feature sequences; Feature difference coefficient calculation and screening module: used to calculate the difference coefficient of feature vectors and screen out suspected matching feature vectors; Matching selection factor calculation and screening module: used to calculate the matching selection factor of the suspected matching feature vector and screen out the matching feature vector; Feature data block similarity evaluation and partitioning module: used to evaluate the similarity of feature data blocks and calculate the partitioning evaluation value to determine the optimal partitioning length; Optimal partition length determination module: used to determine the optimal partition length according to the partition evaluation value; Hash matching retrieval result output module: used to use hash functions for fast matching and output retrieval results.

Images