A SYSTEM AND METHOD FOR ANOMALY DETECTION IN HYPERSPECTRAL IMAGES USING A HYBRID APPROACH.

Abstract

The invention relates to a system and method developed for anomaly detection in hyperspectral data containing hundreds of narrow spectral bands. The invention concerns a hybrid anomaly detection framework aimed at the high-accuracy detection of targets located in scenes with complex backgrounds or camouflage elements (e.g., in fields such as environmental monitoring, agriculture, military reconnaissance, and infrastructure surveillance).

Description

1 TARIFF ANOMALIES IN HYPERSPECTRAL IMAGES USING A HYBRID APPROACH A SYSTEM AND METHOD FOR DETECTION Technical Field to Which the Invention Relates The invention enables anomaly detection in hyperspectral data containing hundreds of narrow spectral bands. It relates to an improved system and method. The invention is based on complex backgrounds or in scenes involving camouflage elements (for example, environmental monitoring, agriculture, military reconnaissance, and high accuracy identification of targets (in areas such as infrastructure monitoring) It discusses a hybrid anomaly detection framework aimed at this purpose. State of the Art 10 Hyperspectral imaging allows a scene to be viewed across different parts of the electromagnetic spectrum. by recording the reflections in the bands, material that the human eye cannot distinguish. It is a powerful remote sensing technique that reveals their differences. In this technique, normal Objects or areas that deviate significantly from background features are called "anomalies". It is called anomaly detection, where such targets are identified without any prior knowledge. 15 It aims to determine this. Various algorithms for hyperspectral anomaly detection exist in the current state of the art. It is used. The best known of these methods is a statistically based approach. It is a Reed-Xiaoli Detector (RXD). RXD detects the spectral signature of a pixel based on its background data. from the statistical model (typically using the mean and covariance matrix 20 (modeled) measures how much it deviates. However, statistical methods like RXD, The high dimensionality inherent in hyperspectral data, high interband range When faced with challenges such as correlation and noise, especially complex or structurally complex issues... It falls short in rich (e.g., urban areas, forested regions) scenes. These methods effectively explore the structural and spatial relationships within the scene. Because it cannot model it, it frequently produces false positives (recognizing background pixels as anomalies). (labeling) produces results. To address these shortcomings, low-rank decomposition was developed. Methods based on this approach have been developed. At the heart of these approaches is a hyperspectral... The image has a background component, which can typically be represented by a low-rank matrix. 30 2 it consists of the sum of an anomaly component that can be represented by a sparse matrix. The assumption lies in Robust Subspace Recovery (RoSuRe) and its variant, Fixed. Methods like Dictionary RoSuRe (FD-RoSuRe) are based on this principle. These methods, hyperspectral data matrix, a low-rank background (L) and sparse anomaly (S) It attempts to model the global structure of the scene by breaking it down into its constituent components. However, these 5 Because the methods only consider global features, pixels are local. They disregard the spectral and spatial context. This is especially true in dense or in complex scenes (for example, airplanes at an airport or in an urban area) (different types of buildings) in accurately defining the boundaries of anomalies and This leads to difficulties in localization. 10 Another approach involves graph-based methods. Laplacian Anomaly Detector (LAD) and... Its variants are LAD-C (Cauchy affinity) and LAD-CS (Cauchy affinity with Spatial) (information), by modeling the similarities between pixels on a graph structure, local It analyzes relationships. These methods treat each pixel as a graph node and It assigns edge weights based on the spectral similarities of pixels. Thus, it makes 15 pixels more similar to their neighbors. They detect pixels that are spectrally significantly different as anomalies. However Because these methods only consider the local context, they fail to grasp the global structure of the scene. They cannot model it. This makes them vulnerable to the overall variations and noise in the background. It makes it sensitive. As a result, it has spectral characteristics similar to the background. Failure to detect anomalies or noisy pixels being mistakenly identified as anomalies 20 Issues such as labeling may arise. In conclusion, the fundamental technical problem in the known state of the art is that the current methods from modeling either only global or only local features This stems from the fact that global methods are structured due to a lack of local context. While failing in the scenes, local methods give way to noise and global background 25 It is sensitive to variations. The combination of these two deficiencies makes it complex and... Accuracy and localization sensitivity of anomaly detection in noisy environments This leads to it being inadequate in that respect. Brief Description and Objectives of the Invention The main purpose of this invention is to enable both global 30-degree anomaly detection in hyperspectral images. a hybrid system and method that models local characteristics in an integrated way to present. 3 Another aim of the invention is to reduce noise and background through global background modeling. The aim is to reduce the effect of variations. For this purpose, hyperspectral data is converted to a low-ranking one. Fixed Dictionary Robust Subspace that decomposes into a background and sparse anomaly matrix. The Recovery (FD-RoSuRe) algorithm is used. This algorithm recovers the background model. By learning directly from the scene itself, it creates a more robust global model. 5 Another aim of the invention is to enable the precise localization of anomalies. The goal is to consider the local spectral and spatial context. To this end, global Laplacian Anomaly on the sparse anomaly matrix obtained from modeling Detection with Cauchy affinity and Spatial information (LAD-CS) algorithm is applied. This algorithm determines both spectral and spatial neighborhood relationships between pixels. It enables more accurate determination of anomaly boundaries. Explanation of the Figures Figure 1: A block diagram showing the general architecture of the system described in the invention. Figure 2: A flowchart showing the working steps of the method described in the invention. Explanation of References in Figures 15 1. Hyperspectral Data Input Unit 2. Robust Subspace Recovery Module 3. Robust Subspace Recovery Module with Fixed Dictionary 4. Low-Rank Background Matrix 5. Sparse Anomaly Matrix 20 6. Laplacian-Based Anomaly Detection Module 7. Spatially Supported Laplacian-Based Anomaly Detection Module 8. Anomaly score map creation unit 1001. Hyperspectral Data Input Unit (1) inputs data for the hyperspectral image. 25 By 1002. Robust Subspace Recovery Module with Fixed Dictionary (3), Low-rank backgrounds are identified by applying a decomposition process on the hyperspectral data matrix. Plan and calculation of sparse components 1003. By the Robust Subspace Recovery Module with Fixed Dictionary (3), parsing As a result of the operation, the low-rank background matrix (4) and the sparse anomaly matrix (5) were obtained. to be done 4 1004. Rare anomalies detected by the Laplacian-Based Anomaly Detection Module (6). Laplacian graph structure based on spectral similarities using the matrix creation and performance of spectral anomaly scoring By the 1005. Spatial Support Laplacian Based Anomaly Detection Module (7), Spectral similarity-based anomaly scoring involves spatial 5-pixel relationships. Inclusion of neighborhood information By the 1006. Spatial-Assisted Laplacian-Based Anomaly Detection Module (7), a composite that reflects both spectral diversity and spatial consistency for each pixel Calculation of anomaly scores 1007. Anomaly Score Map Creation Unit (8) calculated pixel-based 10 Combining anomaly scores to generate the final anomaly score map. Detailed Description of the Invention The invention enables the detection of anomalies in hyperspectral images with high accuracy and precision. It relates to a hybrid system and method that enables localization and identification. The invention The subject matter, system and method, is shown in the block diagram in Figure 1 and the 15 shown in Figure 2. This is explained via a flowchart. The system described in the invention works by feeding a data matrix of a hyperspectral image into the system. This data is provided by the Hyperspectral Data Input Unit (1) and X and is the data matrix shown; where each column of the matrix represents the spectral signature of a pixel. It represents. 20 The first main stage of the system is the global background modeling stage. In this stage, Hyperspectral data matrix X, Robust Subspace Recovery Module with Fixed Dictionary (FD- The module in question is processed by RoSuRe) (3). The module in question is the well-known Robust in the field. It is an improved application of the Subspace Recovery method (RoSuRe) (2). Through this module, the X data matrix 25 X ≈ L + E modeled as follows, a low-rank background matrix (4) (L) with a sparse anomaly The matrix (5) is decomposed into two components (E). The distinguishing feature of this parsing is that the dictionary representing the background is external data. Instead of being extracted from the cluster, it is extracted directly from the internal structure of the hyperspectral data itself. It is the derivation of the scene. This allows for a more accurate depiction of the dominant and homogeneous areas of the scene. is being modeled to be robust against noise and complex background variations. is increased. As a result of decomposition, the low-rank background matrix (L)(4) and sparse 5 An anomaly matrix (E) (5) is obtained (1003). The second main stage of the system is the local anomaly analysis stage. In this stage, infrequent anomaly matrix (5), Spatial Support Laplacian Based Anomaly Detection Module (7) It is provided as input. This module is based on Laplacian-based anomaly detection, a technique known in the field. It is an improved form of the method (6). 10 In the Laplacian-based method, spectral similarities between pixels are used to determine a A graphical structure is created, and anomaly scores are calculated based on this structure. However, this classical approach only takes spectral information into account. The improved module used in the system described in the invention provides, in addition to spectral similarity, It also includes spatial neighborhood relationships in the anomaly assessment. Thus, 15 This prevents individual noise pixels from being falsely labeled as anomalies and ensures accuracy. The spatial continuity of anomalies is determined more accurately at this stage. A combined anomaly score is calculated for each pixel. In the final stage, the calculated pixel-based anomaly scores are combined to identify the anomaly. An anomaly score map is generated by the score map generation unit (8). This 20 The map visually represents the location and relative importance of anomalies throughout the scene. is doing. The (X) notation used in this application represents the hyperspectral image data entered into the system. (L) notation represents the data matrix from the hyperspectral data matrix in question. 25 separated and containing the dominant, homogeneous and background character components of the scene (E) representation of the low-rank background matrix; and (E) representation of the low-rank background component. a potential anomaly that is isolated, spectrally outlier, and sparsely distributed. It represents a sparse anomaly matrix containing pixels. These representations, It has been used to facilitate mathematical and conceptual explanation, 6 any technical element, structural component or independent item included in the scope of the requirements It does not constitute a violation. A hybrid approach to detect anomalies in hyperspectral images. The process consists of the following steps: 1001. Hyperspectral Data Input Unit (1) inputs data for the hyperspectral image. 5 input of the matrix into the system By 1002. Robust Subspace Recovery Module with Fixed Dictionary (3), Low-rank backgrounds are identified by applying a decomposition process on the hyperspectral data matrix. Plan and calculation of sparse components 1003. By the Robust Subspace Recovery Module with Fixed Dictionary (3), parsing 10 As a result of the process, a low-rank background matrix (4) and a sparse anomaly matrix (5) are obtained. to be done 1004. Rare anomalies detected by the Laplacian-Based Anomaly Detection Module (6). Laplacian graph structure based on spectral similarities using the matrix creation and performance of spectral anomaly scoring 15 By the 1005. Spatial Support Laplacian Based Anomaly Detection Module (7), Spectral similarity-based anomaly scoring includes spatial anomaly between pixels. Inclusion of neighborhood information By the 1006. Spatial-Assisted Laplacian-Based Anomaly Detection Module (7), a combined 20 that reflects both spectral difference and spatial consistency for each pixel Calculation of anomaly scores 1007. Pixel-based calculation by the Anomaly Score Map Creation Unit (8). Generating a final anomaly score map by combining the anomaly scores.

Claims

7 REQUESTS 1. A system for detecting anomalies in hyperspectral images; its feature is a server. working on, - at least one hyperspectral system that receives the data matrix of a hyperspectral image into the system. data input unit (1), 5 - On the data matrix provided by the hyperspectral data input unit (1), At least one robust subspace recovery that performs robust subspace recovery. acquisition module (2) - On the data matrix provided by the hyperspectral data input unit (1), the data 10 with a low-rank background component and a sparse anomaly component robust subspace recovery module with at least one fixed dictionary that parses (3), - by robust subspace recovery module (3) with fixed dictionary The dominant element of the scene is obtained as a result of the separation process performed. at least one low-ranking component representing its homogeneous and background character components background matrix (4), 15 - by robust subspace recovery module (3) with fixed dictionary obtained as a result of the separation process and spectrally anomaly pixels, representing potential anomaly pixels with a sparse distribution sparse anomaly matrix (5), - Anomaly scoring based on spectral similarities between pixels 20 at least one Laplacian-based anomaly detection module configured to do so (6), - taking into account the spatial neighborhood relationships between pixels combined for each pixel on the structured and sparse anomaly matrix (5) at least one spatially supported 25 working to calculate an anomaly score Laplacian-based anomaly detection module (7), - by spatially supported Laplacian-based anomaly detection module (7) By combining the calculated pixel-based anomaly scores, across the entire scene a visualizable output showing the location and relative importance of anomalies generating at least one anomaly score map generation unit (8) 30 It includes.

2. It is a method for detecting anomalies in hyperspectral images, and its characteristic feature is; 8 1001. Hyperspectral Data Input Unit (1) of the hyperspectral image importing the data matrix into the system By 1002. Robust Subspace Recovery Module with Fixed Dictionary (3), Low-rank data is extracted by applying a decomposition process on the hyperspectral data matrix. Calculation of background and sparse components 5 By 1003. Robust Subspace Recovery Module with Fixed Dictionary (3), As a result of the decomposition process, the low-rank background matrix (4) and sparse obtaining anomaly matrix (5) 1004. By the Laplacian Based Anomaly Detection Module (6), infrequent Laplacian plot based on spectral similarities using an anomaly matrix 10 creation of the structure and performance of spectral anomaly scoring 1005. Spatial-Assisted Laplacian-Based Anomaly Detection Module (7) by, spectral similarity-based anomaly scoring between pixels Inclusion of spatial neighborhood information 1006. Spatial-Assisted Laplacian-Based Anomaly Detection Module (7) 15 by, spectral difference and spatial consistency together for each pixel. Calculation of combined anomaly scores reflecting 1007. Calculated by the Anomaly Score Map Creation Unit (8) Combining pixel-based anomaly scores to create the final anomaly score map. 20 produced It includes the steps involved in the process.

3. Is there a system for detecting anomalies in hyperspectral images that comply with Claim 1? This feature allows each column to represent the spectral signature of a pixel. Perspectral data input unit providing an organized hyperspectral data matrix. (1) includes. 25 4. Is there a system for detecting anomalies in hyperspectral images that comply with Claim 1? The feature is the dictionary used for the low-rank background matrix (4), an external one. direct hyperspectral data matrix, independent of the reference dataset. robust subspace recovery module with constant dictionary deriving from it (3) It includes. 30 5. Is there a system for detecting anomalies in hyperspectral images that comply with Claim 1? The feature is that the hyperspectral data matrix is ​​sparsely distributed with the low-rank background matrix (4). robust subspace with constant dictionary modeling in the form of sum of anomaly matrix (5) It includes a recovery module (3).

Images