Multi-dimensional knowledge analysis and decision support system based on image recognition

Abstract

The application discloses a multi-dimensional knowledge analysis and decision support system based on image recognition, which comprises image acquisition modules, preprocessing modules, feature extraction modules, knowledge graph construction modules and decision support modules which are connected in sequence, the image acquisition modules and the preprocessing modules realize bidirectional data communication, the output end of the preprocessing modules is connected to the input end of the feature extraction modules, the feature extraction modules are internally provided with feature extraction units and feature calculation units, the output end of the feature extraction modules is connected to the input end of the knowledge graph construction modules, the knowledge graph construction modules contain dynamic updating units, and the knowledge graph construction modules and the decision support modules realize bidirectional data interaction, and the application realizes comprehensive improvement of system performance through multi-dimensional technical innovation, the accuracy of lung nodule benignity and malignancy judgment in the medical image diagnosis scene is significantly improved compared with traditional CNN models, and the recognition rate of the industrial quality inspection system on bearing cracks is simultaneously improved.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and information processing technology, specifically relating to a multi-dimensional knowledge analysis and decision support system based on image recognition. Background Technology

[0002] Image recognition technology has been widely applied in multiple fields. Mainstream solutions are based on convolutional neural network (CNN) architectures. For example, the ResNet series models achieve a Top-5 accuracy of 96.43% on the ImageNet dataset, and the U-Net model generally has a Dice coefficient exceeding 0.85 in medical image segmentation tasks. Knowledge graph technology constructs entity relationship networks using graph databases such as Neo4j, enabling structured knowledge representation in scenarios such as financial risk control and intelligent question answering. However, its application is limited to static data association analysis. For instance, the update cycle of traditional medical knowledge graphs typically exceeds three months, making it difficult to adapt to rapidly evolving clinical research findings. Existing decision support systems mostly adopt a linear process of "data input - rule matching - result output." For example, while IBM Watson Health integrates medical literature and case data, it still faces technical bottlenecks in the multimodal data fusion stage: the accuracy of associating medical image features with electronic medical record text is only 68.3%, and industrial quality inspection systems experience delays exceeding two seconds in real-time anomaly identification on dynamic production lines.

[0003] Existing technologies suffer from structural defects in multimodal data fusion: high-dimensional image features extracted by computer vision models (such as 3D texture features of CT images) and symbolic knowledge from knowledge graphs (such as ICD-10 disease codes) belong to heterogeneous data spaces and lack a unified semantic mapping mechanism. The lack of a dynamic knowledge update mechanism makes it difficult for the system to cope with real-time changing scenarios. For example, in industrial production lines, when new defect samples appear, traditional knowledge graphs require manual annotation before they can be incorporated into the analysis system, with an average lag time of more than 48 hours, resulting in a product false negative rate of approximately 0.3%.

[0004] The limitations of cross-domain decision support capabilities significantly restrict the practicality of systems: existing systems are mostly designed for specific vertical domains, and the reasoning rules of medical diagnostic systems cannot be directly transferred to agricultural pest and disease identification scenarios. In smart city management, the data disconnect between traffic monitoring video analysis systems and public health event early warning systems leads to a 2-3 hour data delay in population flow heat maps and infection risk assessments during epidemics. Technical pain points have a direct negative impact on practical applications: due to the lack of AI-assisted multi-dimensional analysis tools, the missed diagnosis rate of breast cancer ultrasound diagnosis in grassroots hospitals is 23% higher than that in tertiary hospitals; the quality inspection process in manufacturing suffers approximately 210 million yuan in quality losses annually due to single-dimensional visual analysis; in agricultural production, pest and disease monitoring relying on manual inspections results in a delay of more than 72 hours in prevention and control response, causing a 15-20% reduction in yield per mu. The core root cause of these defects lies in the fact that traditional architectures fail to achieve deep coupling between the image perception layer and the knowledge reasoning layer, and lack a dynamically adaptable cross-domain knowledge transfer mechanism.

[0005] Current technologies fail to achieve deep semantic fusion of image features and knowledge graphs, resulting in a semantic gap between visual perception data and structured knowledge. In medical scenarios, the gray-level co-occurrence matrix features of CT images lack an automatic mapping mechanism with entity relationships in tumor pathology knowledge graphs, requiring manual feature annotation and increasing diagnostic time by 40%. Furthermore, the system lacks sufficient multi-dimensional dynamic reasoning capabilities. Traditional systems, relying on predefined rule bases, cannot cope with the real-time evolution of knowledge systems; for example, after updates to industrial quality inspection standards, the system rule adaptation cycle exceeds 72 hours. Real-time decision support also suffers from response latency; in 1080P video stream analysis scenarios, existing systems have an average processing latency of 1.8 seconds, making it difficult to meet the stringent real-time requirements of applications such as traffic monitoring.

[0006] The technical limitations stem from three main aspects: firstly, the generalization ability of traditional convolutional neural network models is limited to specific task domains, resulting in a significant drop in accuracy when transferred to new scenarios; secondly, the knowledge representation methods are simplistic, with RDF triples failing to express the spatial topological relationships and texture features inherent in images; and thirdly, the decision reasoning engine lacks spatiotemporal correlation modeling, making it unable to handle the dynamic correlation between "insect density - crop growth stage - weather conditions" in agricultural pest and disease monitoring. These technical bottlenecks lead to significant performance deficiencies in existing systems across three dimensions: cross-modal knowledge fusion, dynamic knowledge evolution, and real-time decision response. Therefore, this invention provides a multi-dimensional knowledge analysis and decision support system for image recognition to address the problems mentioned in the background. Summary of the Invention

[0007] To address the problems raised in the background, the purpose of this invention is to provide a multi-dimensional knowledge analysis and decision support system based on image recognition, achieving a comprehensive improvement in system performance through multi-dimensional technological innovation. Regarding decision accuracy, the accuracy of determining the benign or malignant nature of lung nodules in medical imaging diagnosis scenarios is significantly improved compared to traditional CNN models, and the recognition rate of bearing cracks in industrial quality inspection systems is simultaneously enhanced.

[0008] To achieve the above-mentioned technical objectives, the technical solution adopted by the present invention is as follows:

[0009] A multi-dimensional knowledge analysis and decision support system based on image recognition includes, in sequence, an image acquisition module, a preprocessing module, a feature extraction module, a knowledge graph construction module, and a decision support module.

[0010] The image acquisition module and the preprocessing module achieve bidirectional data communication. The output of the preprocessing module is connected to the input of the feature extraction module. The feature extraction module has a built-in feature extraction unit and a feature calculation unit. The output of the feature extraction module is connected to the input of the knowledge graph construction module. The knowledge graph construction module includes a dynamic update unit. The knowledge graph construction module and the decision support module perform bidirectional data interaction.

[0011] Further specifying, the image acquisition module includes a visible light sensor unit and an infrared sensor unit. The visible light sensor unit has a resolution of 4K and a frame rate of 30fps, while the infrared sensor unit operates in the wavelength range of 8-14μm. Both acquire data synchronously and image registration is achieved through an FPGA.

[0012] Furthermore, the preprocessing module adopts an FPGA+GPU heterogeneous computing architecture, which includes a Gaussian filtering denoising unit with 3×3 convolution kernels and an adaptive histogram equalization unit, to achieve a processing speed of 0.5ms / frame for images.

[0013] Furthermore, the dynamic update unit performs clustering processing on the target window image to achieve dynamic optimization of knowledge entities.

[0014] Further specifying, the decision support module is equipped with a multi-dimensional decision tensor model D={E,R,T,S}, where E is an entity set, R is a relation matrix, T is a time feature vector, S is a spatial topology matrix, and D is a four-dimensional tensor, and key decision factors are extracted through tensor decomposition algorithm;

[0015] The entity set E contains the core entities related to decision-making, represented as an m×1 vector; the relation matrix R represents the strength of association between entities, represented as an m×m matrix; the time feature vector T represents the dynamic features in the time dimension, represented as a t×1 vector; the spatial topology matrix S represents the spatial positional relationship of entities, represented as an s×s matrix; and the four-dimensional tensor D integrates the above elements into a four-dimensional tensor of m×m×t×s.

[0016] Furthermore, the decision support module also includes a visual interactive interface, capable of displaying knowledge graphs, decision trees, and heatmaps. Figure 3 The decision-making process and results are presented in various formats, allowing users to adjust and optimize decision parameters.

[0017] Furthermore, it also includes an edge computing unit, which is connected to the feature extraction module via a 5G network, enabling distributed image feature extraction and real-time data processing while reducing the computing load on the central server.

[0018] A multi-dimensional knowledge analysis and decision support method based on image recognition includes the following steps:

[0019] S1 Image Data Acquisition: Acquires raw image data of the target scene through the image acquisition module;

[0020] S2 preprocessing: Denoising, enhancing, and normalizing the original image to generate a standard format image tensor;

[0021] S3 Feature Extraction: Extracts multi-dimensional feature vectors from images through feature extraction units and feature calculation units;

[0022] S4 Knowledge Fusion: Links image feature vectors with external knowledge bases to construct a dynamic knowledge network;

[0023] S5 Decision Reasoning: Based on knowledge networks, it performs multi-dimensional reasoning and outputs decision suggestions including confidence scores.

[0024] Further specifying, the knowledge evolution engine in S4 employs incremental learning, using a time decay factor. The formula for achieving timely knowledge management is as follows:

[0025]

[0026] Where λ is the attenuation coefficient and Δt is the time interval.

[0027] Furthermore, the multi-dimensional reasoning in S5 adopts a combination of deep learning and rule-based reasoning. The deep learning model uses an improved ResNet-50 network, and the rule-based reasoning uses production rule representation. The reasoning results of the two are fused through a weighted voting method.

[0028] The beneficial effects of this invention are:

[0029] This system comprises five core modules forming an organic whole: The image acquisition module acquires raw image data using an industrial-grade camera array (including a visible light / infrared dual-spectrum sensor), and its output interacts with the preprocessing module. The preprocessing module employs an FPGA+GPU heterogeneous computing architecture, transmitting processed image data to the feature extraction module. The feature extraction module integrates feature extraction and feature calculation units, connecting with the knowledge graph construction module to synchronously transmit image feature vectors and spatial coordinate information. The knowledge graph construction module uses a distributed graph computing framework, including a dynamic update unit, and communicates bidirectionally with the decision support module via TCP / IP protocol. The decision support module features a multi-dimensional inference engine and a visual interactive interface, receiving the entity association probability matrix output by the knowledge graph module, processing it with a decision algorithm, and generating a structured decision report.

[0030] This invention achieves a comprehensive improvement in system performance through multi-dimensional technological innovation. In terms of decision accuracy, the accuracy of determining the benign or malignant nature of lung nodules in medical imaging diagnosis scenarios is significantly improved compared to traditional CNN models, and the recognition rate of bearing cracks in industrial quality inspection systems is simultaneously enhanced. Breakthrough progress has been made in knowledge update efficiency, with the entity relationship update latency of the dynamic knowledge graph controlled within 87ms; real-time performance is excellent, with a 1080P video stream processing frame rate of 35fps, meeting the real-time requirements of industrial production line scenarios and traffic monitoring scenarios. Attached Figure Description

[0031] The present invention can be further illustrated by the non-limiting embodiments given in the accompanying drawings;

[0032] Figure 1 This is an overall system module diagram of an embodiment of the multi-dimensional knowledge analysis and decision support system based on image recognition of the present invention;

[0033] Figure 2 This is a system module diagram of an embodiment of the multi-dimensional knowledge analysis and decision support system based on image recognition of the present invention;

[0034] Figure 3 This is a flowchart illustrating the steps of an embodiment of the multi-dimensional knowledge analysis and decision support system based on image recognition of the present invention.

[0035] The main component symbols are explained as follows: Image acquisition module 101, preprocessing module 102, feature extraction module 103, knowledge graph construction module 104, decision support module 105, visible light sensor unit 1011, infrared sensor unit 1012, feature extraction unit 1031, feature calculation unit 1032, edge calculation unit 1033, dynamic update unit 1041. Detailed Implementation

[0036] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be further described below with reference to the accompanying drawings and embodiments. The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0037] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0038] Furthermore, the use of terms such as "first" and "second" in this invention is for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of the various embodiments can be combined with each other, but only on the basis of being achievable by those skilled in the art. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0039] like Figure 1 As shown, the multi-dimensional knowledge analysis and decision support system based on image recognition of the present invention includes an image acquisition module 101, a preprocessing module 102, a feature extraction module 103, a knowledge graph construction module 104, and a decision support module 105 connected in sequence.

[0040] The image acquisition module 101 and the preprocessing module 102 achieve bidirectional data communication. The output of the preprocessing module 102 is connected to the input of the feature extraction module 103. The feature extraction module 103 has a built-in feature extraction unit 1031 and a feature calculation unit 1032. The output of the feature extraction module 103 is connected to the input of the knowledge graph construction module 104. The knowledge graph construction module 104 includes a dynamic update unit 1041. The knowledge graph construction module 104 and the decision support module 105 perform bidirectional data interaction.

[0041] In practical applications, the image acquisition module 101 communicates bidirectionally with the preprocessing module 102 via a PCIe 4.0 bus. The output of the preprocessing module 102 is connected to the input of the feature extraction module 103 via a USB 3.2 Gen2 interface. The feature extraction module 103 has a built-in HOG-LBP cascaded feature extraction unit 1031 and a PHOW pyramid feature calculation unit 1032. Its output is connected to the input of the knowledge graph construction module 104 via a PCIe 5.0 high-speed channel. The knowledge graph construction module 104 includes a dynamic update unit 1041 and interacts bidirectionally with the decision support module 105 via TCP / IP protocol. This system can achieve deep semantic fusion of image features and knowledge graphs, and generate a structured decision report containing confidence scores through a multi-dimensional inference engine.

[0042] In the practical application of this embodiment, the image acquisition module 101 includes a visible light sensor unit 1011 and an infrared sensor unit 1012. The visible light sensor unit 1011 has a resolution of 4K and a frame rate of 30fps. The infrared sensor unit 1012 operates in the wavelength range of 8-14μm. Both acquire images synchronously and image registration is achieved through an FPGA.

[0043] In the practical application of this embodiment, the preprocessing module 102 adopts an FPGA+GPU heterogeneous computing architecture, which includes a Gaussian filter denoising unit 1021 with 3×3 convolution kernels and an adaptive histogram equalization unit 1022, to achieve a processing speed of 0.5ms / frame for images.

[0044] In the practical application of this embodiment, the dynamic update unit 1041 performs clustering processing on the target window image to achieve dynamic optimization of knowledge entities.

[0045] In the practical application of this embodiment, the decision support module 105 is equipped with a multi-dimensional decision tensor model D={E,R,T,S}, where E is an entity set, R is a relation matrix, T is a time feature vector, S is a spatial topology matrix, and D is a four-dimensional tensor. Key decision factors are extracted through tensor decomposition algorithm.

[0046] The entity set E contains the core entities related to decision-making, represented as an m×1 vector; the relation matrix R represents the strength of association between entities, represented as an m×m matrix; the time feature vector T represents the dynamic features in the time dimension, represented as a t×1 vector; the spatial topology matrix S represents the spatial positional relationship of entities, represented as an s×s matrix; and the four-dimensional tensor D integrates the above elements into a four-dimensional tensor of m×m×t×s.

[0047] In practical applications of this embodiment, the decision support module 105 further includes a visual interactive interface, which can display knowledge graphs, decision trees, and heatmaps. Figure 3 The decision-making process and results are presented in various formats, allowing users to adjust and optimize decision parameters.

[0048] In practical applications of this embodiment, an edge computing unit 1033 is also included. The edge computing unit 1033 is connected to the feature extraction module 103 via a 5G network, which enables distributed image feature extraction and real-time data processing while reducing the computing load on the central server.

[0049] A multi-dimensional knowledge analysis and decision support method based on image recognition includes the following steps:

[0050] S1 Image Data Acquisition: Acquire raw image data of the target scene through the image acquisition module 101;

[0051] S2 preprocessing: Denoising, enhancing, and normalizing the original image to generate a standard format image tensor;

[0052] S3 Feature Extraction: Extracts multi-dimensional feature vectors of the image through feature extraction unit 1031 and feature calculation unit 1032;

[0053] S4 Knowledge Fusion: Links image feature vectors with external knowledge bases to construct a dynamic knowledge network;

[0054] S5 Decision Reasoning: Based on knowledge networks, it performs multi-dimensional reasoning and outputs decision suggestions including confidence scores.

[0055] In the practical application of this embodiment, step S4 employs an incremental learning knowledge evolution engine, using a time decay factor. The formula for achieving timely knowledge management is as follows:

[0056]

[0057] Where λ is the attenuation coefficient and Δt is the time interval.

[0058] In the practical application of this embodiment, the multi-dimensional reasoning in S5 adopts a combination of deep learning and rule-based reasoning. The deep learning model adopts an improved ResNet-50 network, and the rule-based reasoning adopts a production rule representation. The reasoning results of the two are fused through a weighted voting method.

[0059] The working principle of this system is as follows:

[0060] The system workflow follows a closed-loop mechanism of "data-driven - knowledge fusion - intelligent decision-making." The startup phase first completes environment configuration: loading the domain knowledge base, initializing model parameters, and configuring hardware interfaces. After entering the runtime phase, the following processing steps are executed:

[0061] In the multimodal sensing stage, the image acquisition module 101 captures visible light / infrared images at a preset frame rate (e.g., 1 fps for medical scenarios, 30 fps for industrial scenarios), and synchronizes multi-source data through timestamps. The raw data is transmitted to the preprocessing module 102 via the GigE interface. The FPGA chip in the preprocessing module 102 performs denoising, enhancement, and resizing operations in parallel, outputting a standardized image tensor stored in the DRAM buffer. The triggering mechanism supports both automatic acquisition and external triggering modes. Abnormal images (overexposed / underexposed) are automatically marked and reacquired through brightness histogram analysis.

[0062] In the feature parsing stage, the preprocessed image tensors are fed into the GPU-accelerated feature extraction module 103 for pipelined processing. The HOG unit built into the feature extraction module 103 computes the gradient direction histogram in parallel using the CUDA kernel function, the LBP unit built into the feature extraction module 103 employs a parallel texture feature extraction algorithm, and the PHOW unit built into the feature extraction module 103 detects feature points in multi-scale space. After dynamic weighted fusion of the three-level features, the dimensionality is reduced to 256-dimensional feature vectors using PCA, and spatial coordinate information (x, y, θ) is added to form a feature package. The feature extraction process adopts an incremental learning mechanism, updating the feature template library every 1000 images processed.

[0063] During the knowledge construction phase, feature packages are transmitted to the knowledge graph server via TCP / IP protocol. The BERT entity extraction unit identifies visual entities in the images and establishes associations with external knowledge bases through entity links. The TransE inference unit calculates potential relationships between entities (such as "located in" or "cause"), and the dynamic update unit 1041 adjusts entity weights according to time decay rules. The completed knowledge network is stored in a Neo4j cluster, supporting both real-time query and batch analysis access modes. Knowledge conflict detection is achieved through confidence comparison; a manual review process is triggered when the confidence difference between two relationships is less than 0.2.

[0064] During the decision generation phase, the decision support module periodically (e.g., every 30 seconds in medical scenarios, every 100 milliseconds in industrial scenarios) pulls entity relationship data from the knowledge graph. The multi-dimensional inference engine performs deep inference and rule-based inference in parallel. The fused decision results are formatted as JSON and pushed to the front-end interface via WebSocket, while simultaneously being stored in a relational database (PostgreSQL) to form a decision log. Abnormal decisions (confidence < 0.7) automatically trigger secondary inference, increasing the depth of feature extraction and the range of rule matching.

[0065] During the feedback optimization phase, the system collects user feedback on the decision results (acceptance / rejection / correction) and updates the inference model parameters through reinforcement learning: when a user accepts a decision, the weight of the corresponding inference path is increased; when a user corrects a decision, backpropagation is performed to adjust the feature weights. A weekly performance report is generated, including key metrics such as accuracy, recall, and F1 score. When these metrics fall below a threshold (0.85 in the medical field, 0.90 in the industrial field), the model retraining process is automatically initiated. The system supports OTA upgrades, and a canary release strategy is used when updating modules to ensure service continuity.

[0066] The key node processing logic is implemented using a state machine model: image acquisition states include four types: idle, acquisition, transmission, and error; feature extraction states include four types: waiting for data, processing, completed, and failed; knowledge construction states include four types: entity extraction, relation reasoning, update, and conflict detection. State transitions are achieved through an event-driven mechanism, such as the "feature extraction completed" event triggering the knowledge construction process, and the "decision timeout" event triggering degradation processing (using a simplified model). The system avoids concurrent conflicts through a distributed lock mechanism and uses a circuit breaker mechanism to prevent cascading failures, ensuring that a single-point module failure does not affect the overall system operation.

[0067] Example 1:

[0068] Medical Imaging Tumor Diagnosis System: In clinical application at the radiology department of a tertiary hospital, this system is deployed on an image analysis workstation to achieve intelligent processing of the entire lung cancer diagnosis process. The system receives chest CT image data in DICOM format (0.625mm slice thickness, 512×512 matrix). First, it performs automatic lung parenchyma segmentation using an improved 3D U-Net model, achieving a segmentation accuracy of 0.983 (Dice coefficient). The nodule detection module employs a multi-scale region generation strategy, achieving a 96.7% detection rate for intrapulmonary nodules with a diameter of 5-30mm, with false positives controlled below 2.1 per case.

[0069] The feature extraction stage simultaneously calculates 12 quantitative parameters: three-dimensional diameter (major diameter / minor diameter / volume), morphological features (lobulation / spiculation / pleural traction), density features (average CT value / calcification ratio), and texture features (entropy / energy / contrast). The knowledge graph module links these features with the hospital's electronic medical record system (HIS) and pathology database to construct a four-dimensional knowledge network containing "patient basic information, imaging features, clinical diagnosis, and pathological results." Taking the CT image of a 65-year-old male patient (with a 30-year smoking history) as an example, the system automatically identifies an 8mm mixed-density nodule in the right upper lung apex segment. Through knowledge graph matching, it matches the feature combination of "smoking history, pleural traction sign, and cavitation sign," triggering the malignant tumor inference rule.

[0070] The multi-dimensional inference engine outputs benign / malignant results: the deep learning channel gives a malignancy probability of 0.91, while the rule-based inference channel, through a rule chain of "IF nodule edge spiculation AND volume > 500 mm³ AND SUVmax > 2.5 THEN Malignant tumor (confidence 0.93)," ultimately achieves a malignancy confidence of 0.92. The entire analysis process takes 4.7 seconds, which is dozens of times more efficient than traditional manual image reading (average 6.5 minutes). Furthermore, in 200 clinical trials, the system achieved a diagnostic accuracy of 97.2%.

[0071] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. A multi-dimensional knowledge analysis and decision support system based on image recognition, characterized in that: It includes an image acquisition module (101), a preprocessing module (102), a feature extraction module (103), a knowledge graph construction module (104), and a decision support module (105) connected in sequence. The image acquisition module (101) and the preprocessing module (102) achieve bidirectional data communication. The output of the preprocessing module (102) is connected to the input of the feature extraction module (103). The feature extraction module (103) has a built-in feature extraction unit (1031) and a feature calculation unit (1032). The output of the feature extraction module (103) is connected to the input of the knowledge graph construction module (104). The knowledge graph construction module (104) includes a dynamic update unit (1041). The knowledge graph construction module (104) and the decision support module (105) perform bidirectional data interaction.

2. The multi-dimensional knowledge analysis and decision support system based on image recognition according to claim 1, characterized in that: The image acquisition module (101) includes a visible light sensor unit (1011) and an infrared sensor unit (1012). The visible light sensor unit (1011) has a resolution of 4K and a frame rate of 30fps. The infrared sensor unit (1012) operates in the wavelength range of 8-14μm. The two sensors acquire images synchronously and image registration is achieved through an FPGA.

3. The multi-dimensional knowledge analysis and decision support system based on image recognition according to claim 1, characterized in that: The preprocessing module (102) adopts an FPGA+GPU heterogeneous computing architecture, which includes a Gaussian filtering denoising unit (1021) with 3×3 convolution kernels and an adaptive histogram equalization unit (1022) to achieve a processing speed of 0.5ms / frame for images.

4. The multi-dimensional knowledge analysis and decision support system based on image recognition according to claim 1, characterized in that: The dynamic update unit (1041) performs clustering processing on the target window image to achieve dynamic optimization of knowledge entities.

5. The multi-dimensional knowledge analysis and decision support system based on image recognition according to claim 1, characterized in that: The decision support module (105) is equipped with a multi-dimensional decision tensor model D={E,R,T,S}, where E is an entity set, R is a relation matrix, T is a time feature vector, S is a spatial topology matrix, and D is a four-dimensional tensor. Key decision factors are extracted through tensor decomposition algorithm. The entity set E contains the core entities related to decision-making, represented as an m×1 vector; the relation matrix R represents the strength of association between entities, represented as an m×m matrix; the time feature vector T represents the dynamic features in the time dimension, represented as a t×1 vector; the spatial topology matrix S represents the spatial positional relationship of entities, represented as an s×s matrix; and the four-dimensional tensor D integrates the above elements into a four-dimensional tensor of m×m×t×s.

6. The multi-dimensional knowledge analysis and decision support system based on image recognition according to claim 1, characterized in that: The decision support module (105) also includes a visual interactive interface that can display the decision process and results in three forms: knowledge graph, decision tree and heat map, and supports users to adjust and optimize decision parameters.

7. The multi-dimensional knowledge analysis and decision support system based on image recognition according to claim 1, characterized in that: It also includes an edge computing unit (1033), which is connected to the feature extraction module (103) via a 5G network, enabling distributed image feature extraction and real-time data processing while reducing the computing load on the central server.

8. A multi-dimensional knowledge analysis and decision support method based on image recognition, characterized in that, Includes the following steps: S1 Image Data Acquisition: Acquire raw image data of the target scene through the image acquisition module (101); S2 preprocessing: Denoising, enhancing, and normalizing the original image to generate a standard format image tensor; S3 Feature Extraction: Extracts multi-dimensional feature vectors of the image through the feature extraction unit (1031) and the feature calculation unit (1032); S4 Knowledge Fusion: Links image feature vectors with external knowledge bases to construct a dynamic knowledge network; S5 Decision Reasoning: Based on knowledge networks, it performs multi-dimensional reasoning and outputs decision suggestions including confidence scores.

9. The multi-dimensional knowledge analysis and decision support method based on image recognition according to claim 8, characterized in that: S4 employs an incremental learning knowledge evolution engine, which utilizes a time decay factor. The formula for achieving timely knowledge management is as follows: Where λ is the attenuation coefficient and Δt is the time interval.

10. The multi-dimensional knowledge analysis and decision support method based on image recognition according to claim 8, characterized in that: The multi-dimensional reasoning in S5 adopts a combination of deep learning and rule-based reasoning. The deep learning model uses an improved ResNet-50 network, and the rule-based reasoning uses production rule representation. The reasoning results of the two are fused through weighted voting.

Images