Sparse representation method and system based on dynamic multi-scale dictionary generation
Through the dynamic multi-scale dictionary generation method, the limitations of traditional sparse dictionary in complex point cloud and multimodal data processing are solved, and more efficient sparse encoding and feature extraction are achieved, which improves recognition accuracy and noise resistance.
Patent Information
- Application Number
- CN202510639877.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-19
- Publication Date
- 2025-08-19
AI Technical Summary
When traditional sparse dictionaries deal with complex radar point cloud data and multimodal data, it is difficult to accurately capture local geometric mutations and cross-modal features, resulting in inaccurate identification of obstacles and noise residues, and cannot effectively adapt to the heterogeneous characteristics of multimodal data.
The dynamic multi-scale dictionary generation method is adopted to extract point cloud features through local voxel block division and K-SVD algorithm, combine DCT and random noise blocks to construct image dictionary, and use multi-scale region division and dynamic atomic optimization strategies to eliminate redundant atoms, enhance key feature primitives, and achieve dynamic matching of dictionary structure and data characteristics.
It significantly improves the sparse encoding capability of complex geometric structures and high-frequency textures, improves the feature extraction and representation accuracy of point clouds and image data, enhances the detection ability against attacks, and adapts to the feature representation of multimodal data.
Smart Images
Figure CN120510399A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of information technology and relates to a sparse representation method and system based on dynamic multi-scale dictionary generation. Background Art
[0002] Sparse dictionaries play a crucial role in many cutting-edge fields, including signal processing, image processing, and machine learning. Traditional sparse dictionaries are essentially a set of carefully designed basis functions, like precise keys that unlock the door to sparse representation of signals or images. The core idea of sparse representation is to use a relatively small number of basis function combinations to approximate the original signal or image, significantly reducing data complexity while preserving critical information.
[0003] In the field of signal processing, sparse dictionaries play an irreplaceable role. For example, by selecting an appropriate sparse dictionary, audio signals can be converted from the time domain to the sparse domain. This process effectively removes redundant information from the signal, retaining only the components that contribute most to the audio's essential characteristics. This not only significantly reduces the storage space for audio data but also provides a more concise and efficient data representation for subsequent audio processing tasks, such as audio classification and speech recognition. Sparse dictionaries also demonstrate powerful advantages in image processing. Fixed-structure sparse dictionaries, such as the common DCT (Discrete Cosine Transform) and wavelet bases, have become effective tools for image compression and denoising. In image compression, these dictionaries are used to perform sparse decomposition of an image, retaining only the important sparse coefficients and discarding those that have little impact on the overall visual quality, thereby achieving efficient image compression. For example, the JPEG image compression standard makes extensive use of the DCT transform, significantly reducing image file size while maintaining acceptable image quality. In image denoising, by thresholding the sparse coefficients, we can accurately distinguish between signal components and noise components, retaining the main signal components while suppressing noise, effectively improving the clarity and quality of the image.
[0004] Traditional sparse dictionaries are widely used in these fields due to their many advantageous properties. These sparse dictionary-based transformations are mostly linear, a property that greatly simplifies mathematical processing and analysis. In practical applications, linear transformations have relatively low computational complexity, enabling rapid signal or image processing, meeting the demands of real-time applications. Furthermore, in image processing, the use of fixed dictionaries provides robust feature representation. Regardless of the degree of noise or slight deformation present in the image, fixed dictionaries maintain a certain degree of ability to extract key image features, providing a reliable foundation for subsequent image analysis and understanding.
[0005] However, with the rapid advancement of technology and the increasing complexity of application scenarios, the limitations of traditional sparse dictionaries have gradually become apparent. In the specialized field of radar point cloud modalities, the fixed atomic structure of traditional sparse dictionaries has become a key factor limiting their performance. Radar point cloud data contains rich three-dimensional spatial information, with complex and variable local geometric structures and a large number of geometric mutations. The fixed atoms of traditional sparse dictionaries are unable to accurately capture these local geometric mutations, resulting in distortion or loss of detail during data reconstruction. For example, in the field of autonomous driving, radar point cloud data is crucial for accurately identifying obstacles in the surrounding environment. Using traditional sparse dictionaries for point cloud data processing can lead to inaccurate obstacle recognition due to the inability to accurately represent local geometric features, thus compromising the safety of autonomous driving.
[0006] In depth image processing, a single dictionary also exhibits significant shortcomings. Depth images record distance information on the surface of objects, and their data characteristics differ from those of traditional two-dimensional images. When processing depth images, a single dictionary often struggles to simultaneously account for both smooth areas and edge details. This can easily lead to oversmoothing during denoising, blurring image edges and texture information. Alternatively, noise cannot be effectively suppressed while retaining detail, resulting in residual noise.
[0007] What is more complicated is that in today's context of the increasing popularity of multimodal data, traditional sparse dictionaries face greater challenges. Multimodal data has heterogeneous characteristics, and data of different modalities have significant differences in data structure, feature distribution, etc. It is difficult for a general dictionary to find a balance between representation efficiency and cross-domain consistency. In cross-modal detection tasks, such as simultaneously processing data of multiple modalities such as images, text, and speech, traditional sparse dictionary construction methods are difficult to adapt to the characteristics of data of different modalities and cannot effectively extract shared features and complementary information between modalities, thereby limiting its application effect in cross-modal detection tasks. In summary, although traditional sparse dictionaries have achieved remarkable application results in the past, their limitations are becoming increasingly apparent when faced with complex data characteristics and multimodal data processing requirements. New theories and methods are urgently needed to break through these bottlenecks. Summary of the Invention
[0008] In light of this, the present invention aims to provide a sparse representation method and system based on dynamic multi-scale dictionary generation. This method and system automatically remove redundant atoms and enhance key feature primitives during training, dynamically matching the dictionary structure with data characteristics. Through multi-stage iterative optimization, this solution significantly improves the sparse encoding capabilities for complex geometric structures and high-frequency textures.
[0009] In order to achieve the above object, the present invention provides the following technical solutions:
[0010] A sparse representation method based on dynamic multi-scale dictionary generation, the method specifically comprising the following steps:
[0011] S1. Multimodal dictionary initialization, including:
[0012] S11. For point cloud data, local voxel blocks are used for structured processing, and local geometric feature primitives are extracted from the voxel blocks through the K-SVD algorithm to generate an initial dictionary;
[0013] S12, constructing an initial dictionary for the image data by combining discrete cosine transform (DCT) basis functions with random noise blocks;
[0014] S2. Multi-scale region division and dynamic optimization, including:
[0015] S21. In the point cloud data, based on local curvature analysis, the low curvature, medium curvature and high curvature regions are divided and atoms of different densities are assigned for sparse representation.
[0016] S22. In the image data, the frequency domain energy distribution is analyzed by Haar wavelet transform, low-frequency smooth, medium-frequency and high-frequency texture areas are divided, and different atomic combinations are adapted;
[0017] S3. Using a dynamic atom optimization strategy, the dictionary structure is adjusted through a hierarchical pruning-refinement mechanism. This includes regularly screening and removing redundant atoms whose usage rate is below a set threshold; and removing redundant atom pairs whose inter-atom correlation coefficient is higher than a preset value to improve the orthogonality of the dictionary.
[0018] S4. Cross-modal joint coding: Align the features of the point cloud and the optimized dictionary of the image to build a shared atomic library; constrain the atomic combination consistency of multimodal signals through sparse coding to detect cross-modal correlation anomalies.
[0019] Furthermore, in step S11, the size of the voxel block is divided into 8×8×8, and the plane, edge and corner features of the point cloud are captured by iteratively updating sparse coding and dictionary atoms.
[0020] Furthermore, in step S21, it specifically includes calculating the k-nearest neighbor covariance matrix of each point, and obtaining the principal curvature value through eigenvalue decomposition; dividing the region according to the curvature range: curvature <0.1 is a low curvature region, and sparse atom allocation is adopted (for example, 3-5 atoms are allocated to each voxel block); 0.1≤curvature≤0.3 is a medium curvature region, and an appropriate number of atoms are used to capture the local geometric structure (for example, 6-8 atoms are allocated); curvature >0.3 is a high curvature region, and more refined atoms are used for representation (for example, 10-12 atoms are allocated) to ensure that the geometric information of the point cloud is not lost.
[0021] Further, in step S22, the frequency domain energy analysis based on Haar wavelet transform divides the regions specifically into the following: high-frequency energy <5 is a low-frequency smooth region, which is encoded using pure DCT-based atoms; 5≤high-frequency energy≤20 is a medium-frequency region, which is encoded using a combination of DCT-based atoms and random noise blocks; high-frequency energy >20 is a high-frequency texture region, which is encoded using a combination of random noise blocks and high-frequency atoms.
[0022] Furthermore, in step S3, during the hierarchical pruning process, the atom usage is regularly counted, atoms with a usage rate lower than 5% are deleted, and redundant atom pairs with an inter-atom correlation coefficient higher than 0.8 are eliminated.
[0023] The present invention also provides a sparse representation system based on dynamic multi-scale dictionary generation, which includes an initialization module, a feature perception module, an optimization module, an encoding module, and an anomaly detection module;
[0024] The initialization module is configured to generate an adaptive initial dictionary according to the data type, including a point cloud voxel division unit and an image mixing unit;
[0025] The feature perception module includes a curvature analysis unit (for point clouds) and a wavelet frequency domain division unit (for images), and outputs region division results and atom allocation strategies.
[0026] The optimization module is configured to perform hierarchical pruning, segmentation enhancement, and cross-modal alignment, wherein the pruning unit uses a dynamic threshold (usage rate <5%, correlation coefficient >0.8) to screen atoms;
[0027] The encoding module performs sparse encoding on the input signal based on the optimized dictionary, and outputs sparse coefficients and a reconstructed signal;
[0028] The anomaly detection module detects physical adversarial attacks or data anomalies through cross-modal atomic combination consistency analysis.
[0029] The present invention also provides a computer-readable storage medium storing a computer program, which implements the above method when executed by a processor.
[0030] The beneficial effects of the present invention are:
[0031] The technical solution provided by this invention employs different optimization measures for point cloud and image modalities. In point cloud data, anisotropic Gaussian kernel atoms are introduced to improve the representation accuracy of edge regions in areas with sudden changes in curvature. In image data, high-frequency texture atoms are synthesized using a generative adversarial network to enhance the ability to represent details. Through this dynamic optimization mechanism, the dictionary can continuously adapt to changes in data distribution, thereby achieving better sparse coding results.
[0032] Other advantages, objects, and features of the present invention will be described in part in the following description and, in part, will be apparent to those skilled in the art upon examination of the following description or may be learned from practice of the present invention. The objects and other advantages of the present invention may be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS
[0033] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below with reference to the accompanying drawings, in which:
[0034] Figure 1 Sparse representation graph generated for dynamic multi-scale dictionary;
[0035] Figure 2 This is a diagram of the division and calculation logic of point cloud data. DETAILED DESCRIPTION
[0036] The technical solution of the present invention is described in detail below with reference to the accompanying drawings.
[0037] This paper proposes a sparse representation method and system based on dynamic multi-scale dictionary generation. The core idea is to automatically remove redundant atoms and enhance key feature primitives during training to achieve dynamic matching between dictionary structure and data characteristics. This technology significantly improves the sparse coding capability of complex geometric structures and high-frequency textures through multi-stage iterative optimization. Figure 1 The sparse representation graph generated by the dynamic multi-scale dictionary of the present invention.
[0038] The cross-modal joint dictionary learning scheme aims to establish a shared "feature vocabulary" for multimodal data, allowing signals from multiple different modalities to be expressed using a combination of the same set of basic elements (called "atoms") when encoding. This consistent encoding mechanism can effectively capture the inherent correlations in multimodal data, while physical adversarial attacks will destroy these correlations and thus be detected through encoding anomalies. The method provided by the present invention specifically includes the following steps:
[0039] Dictionary initialization:
[0040] During the dictionary initialization phase, different strategies are employed across multiple modalities to ensure dictionary diversity and adaptability, enabling them to effectively adapt to the data characteristics of their respective modalities and enhance subsequent feature extraction and representation. In the point cloud modality, local voxel block partitioning (e.g., 8×8×8 voxel blocks) is proposed to structure the point cloud data to reduce data sparsity and enhance local geometric consistency. Furthermore, the K-SVD algorithm is applied to generate the initial dictionary. This method adaptively extracts representative atomic primitives from the data within an unsupervised learning framework, capturing the geometric structural features of the point cloud, such as planes, edges, and corners. Compared to traditional PCA or K-means clustering methods, K-SVD, through iterative dictionary updates and sparse coding, can more accurately learn local feature patterns in point clouds, improving the representation capabilities of point cloud data and enhancing its adaptability to different scales and complex scenes. In the image modality, the study combines DCT (discrete cosine transform) basis functions with random noise blocks to construct the dictionary, balancing low-frequency global information with high-frequency texture details. The basis functions of the DCT transform can effectively extract the low-frequency components of the image and provide a robust representation of the overall structure, while the introduction of random noise blocks can increase the diversity of the dictionary, enabling it to better adapt to changes in local texture. Compared with dictionary construction methods that only use the DCT transform or learning-based methods, this hybrid strategy can ensure the adaptability of the dictionary at different scales and improve its sparse representation capabilities in complex visual scenes. In addition, this strategy also has advantages in computational efficiency and sparsity control, which can enhance the model's ability to capture image details while maintaining efficient coding.
[0041] Multi-scale feature perception:
[0042] In order to further improve the adaptability of the dictionary, the present invention proposes a multi-scale feature perception method, which performs feature region division for point cloud and image data respectively, thereby realizing regional adaptive dictionary optimization and improving the sparse representation capability of different modal data. In the point cloud modality, the study is based on principal curvature analysis, and uses PCA covariance matrix eigenvalue decomposition to calculate the local curvature of the point cloud, and divides the geometric complexity area accordingly. Specifically, the calculation process of the local curvature is as follows: First, the covariance matrix is calculated for the k nearest neighbors (k-NN) of each point, and its principal curvature information is obtained by eigenvalue decomposition. According to the size of the curvature, the point cloud can be divided into the following three types of areas:
[0043] 1) Low curvature regions (curvature < 0.1): These regions mainly correspond to flat surfaces, such as the ground and walls. Their geometric features are relatively simple and can be encoded with fewer atoms to reduce computational burden and improve storage efficiency.
[0044] 2) Medium curvature regions (0.1 ≤ curvature ≤ 0.3): These regions typically contain slowly changing surfaces, such as cylinders or arcs, with features between flat and high curvature regions, requiring a moderate number of atoms to capture the local geometry;
[0045] 3) High curvature areas (curvature > 0.3): These areas contain complex structures such as edges, corners, and sudden curves. They are rich in information and have high recognition. Therefore, more refined atoms are required to represent them to ensure that the geometric information of the point cloud is not lost.
[0046] The overall division logic is as follows Figure 2 As shown in Figure 3, this adaptive dictionary strategy enables point clouds to adopt adaptive atomic allocation in areas of different complexity, improving the representation efficiency of the dictionary while avoiding the storage and calculation of redundant information.
[0047] In the image modality, the study used Haar wavelet transform to analyze the frequency domain energy distribution of the image, thereby dividing the low-frequency smooth area and the high-frequency texture area. Haar wavelet transform is a fast and efficient multi-scale analysis method that can decompose the image into frequency components of different scales and directions. Based on the calculation of high-frequency energy, the study divided the image into:
[0048] 1) Low-frequency smooth regions (high-frequency energy < 5): These regions are primarily composed of slowly changing grayscale or color distributions, such as large smooth areas like the sky and walls. In these regions, the dictionary primarily relies on the DCT low-frequency basis functions for modeling to ensure a stable global representation.
[0049] 2) Medium frequency region (5≤high frequency energy≤20): This region usually contains soft textures or edges, such as grass and leaves. Its local features are rich but change slowly, and an appropriate number of atoms are required to express them to ensure the integrity of local information.
[0050] 3) High-frequency texture regions (high-frequency energy > 20): These regions typically contain fine textures, sharp edges, or complex patterns, such as text, object edges, and noise interference. In these regions, the study introduced random noise blocks combined with high-frequency atoms to enhance the adaptability of the dictionary, thereby improving the representation of complex structures.
[0051] Through this region-based adaptive optimization strategy, the constructed dictionary can be dynamically adjusted in different image regions, so that smooth areas maintain a stable sparse expression, while high-frequency areas can more fully capture complex texture details, thereby improving the overall sparse representation effect.
[0052] Dynamic atomic optimization strategy:
[0053] In terms of dictionary optimization, the present invention adopts a dynamic atomic optimization strategy, that is, through pruning and subdivision mechanisms, the atomic structure of the dictionary is continuously adjusted to improve coding efficiency and reduce redundancy. In the pruning strategy, a method of regularly screening and removing redundant atoms is adopted to improve the compactness of the dictionary and reduce the ineffective computational burden. During the use of the dictionary, the study regularly counts the frequency of atomic usage and removes atoms with a usage rate of less than 5%. These atoms often contribute little and may become redundant or no longer applicable to new samples after changes in the data distribution. Therefore, removing them can avoid wasting dictionary resources and accelerate the convergence of sparse coding. When updating the dictionary, the mutual correlation coefficient between atoms is calculated, and atom pairs with a correlation coefficient greater than 0.8 are removed. Highly correlated atoms often mean that there are redundant primitives in the dictionary, causing the encoding process to produce unnecessary repeated representations, thereby affecting the optimization effect of sparsity. Removing these atoms can improve the orthogonality of the dictionary, making its distribution in high-dimensional space more uniform, thereby improving coding efficiency.
[0054] The method of the present invention is compared and analyzed with traditional fixed dictionary methods (K-SVD, DCT), classic sparse representation methods (OMP, LASSO) and recent dynamic dictionary methods. Compared with existing methods, the method of the present invention has better applicability and advantages in different tasks.
[0055] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention can be modified without departing from the purpose and scope of the technical solutions, which should all be included in the scope of the claims of the present invention.
Claims
1. A sparse representation method based on dynamic multi-scale dictionary generation, characterized by: The method specifically comprises the following steps: S1. Multimodal dictionary initialization, including: S11. For point cloud data, local voxel blocks are used for structured processing, and local geometric feature primitives are extracted from the voxel blocks through the K-SVD algorithm to generate an initial dictionary; S12, constructing an initial dictionary for the image data by combining discrete cosine transform (DCT) basis functions with random noise blocks; S2. Multi-scale region division and dynamic optimization, including: S21. In the point cloud data, based on local curvature analysis, the low curvature, medium curvature and high curvature regions are divided and atoms of different densities are assigned for sparse representation. S22. In the image data, the frequency domain energy distribution is analyzed by Haar wavelet transform, low-frequency smooth, medium-frequency and high-frequency texture areas are divided, and different atomic combinations are adapted; S3. Using a dynamic atom optimization strategy, the dictionary structure is adjusted through a hierarchical pruning-refinement mechanism. This includes regularly screening and removing redundant atoms whose usage rate is below a set threshold; and removing redundant atom pairs whose inter-atom correlation coefficient is higher than a preset value to improve the orthogonality of the dictionary. S4. Cross-modal joint coding: Align the features of the point cloud and the optimized dictionary of the image to build a shared atomic library; constrain the atomic combination consistency of multimodal signals through sparse coding to detect cross-modal correlation anomalies.
2. The sparse representation method based on dynamic multi-scale dictionary generation according to claim 1, characterized in that: In step S11, the size of the voxel block is divided into 8×8×8, and the plane, edge and corner features of the point cloud are captured by iteratively updating sparse coding and dictionary atoms.
3. The sparse representation method based on dynamic multi-scale dictionary generation according to claim 2, characterized in that: In step S21, the k-nearest neighbor covariance matrix of each point is calculated, and the principal curvature value is obtained by eigenvalue decomposition; the region is divided according to the curvature range: the curvature <0.1 is a low curvature region, and sparse atom allocation is adopted; 0.1≤curvature≤0.3 is a medium curvature area, and an appropriate number of atoms are used to capture the local geometric structure; curvature>0.3 is a high curvature area, and more refined atoms are used to represent it to ensure that the geometric information of the point cloud is not lost.
4. The sparse representation method based on dynamic multi-scale dictionary generation according to claim 3, characterized in that: In step S22, the frequency domain energy analysis based on Haar wavelet transform divides the regions into the following specific areas: high-frequency energy <5 is a low-frequency smooth region, which is encoded using pure DCT-based atoms; 5≤high-frequency energy≤20 is a medium-frequency region, which is encoded using a combination of DCT-based atoms and random noise blocks; high-frequency energy >20 is a high-frequency texture region, which is encoded using a combination of random noise blocks and high-frequency atoms.
5. The sparse representation method based on dynamic multi-scale dictionary generation according to claim 4, characterized in that: In step S3, during the hierarchical pruning process, the atom usage is regularly counted, atoms with a usage rate lower than 5% are deleted, and redundant atom pairs with an inter-atom correlation coefficient higher than 0.8 are eliminated.
6. A sparse representation system based on dynamic multi-scale dictionary generation, characterized by: The system includes an initialization module, a feature perception module, an optimization module, an encoding module, and an anomaly detection module; The initialization module is configured to generate an adaptive initial dictionary according to the data type, including a point cloud voxel division unit and an image mixing unit; The feature perception module includes a curvature analysis unit and a wavelet frequency domain division unit, and outputs the region division result and the atom allocation strategy; The optimization module is configured to perform hierarchical pruning, segmentation enhancement, and cross-modal alignment, wherein the pruning unit uses a dynamic threshold to screen atoms; The encoding module performs sparse encoding on the input signal based on the optimized dictionary, and outputs sparse coefficients and a reconstructed signal; The anomaly detection module detects physical adversarial attacks or data anomalies through cross-modal atomic combination consistency analysis.
7. A computer-readable storage medium storing a computer program, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 5 is implemented.