Artificial intelligence image analysis
By combining dynamic noise filtering and intelligent region segmentation with multi-dimensional feature extraction and parameter library updates, the robustness and adaptability issues of existing image analysis methods in complex scenarios are solved, achieving high-precision and reliable image recognition and data management.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- 姚子晨
- Filing Date
- 2026-02-27
- Publication Date
- 2026-06-30
AI Technical Summary
Existing image analysis methods lack robustness in complex scenarios, struggle to adapt to changes in image quality, suffer from incomplete feature extraction, lack dynamic parameter optimization, and lack effective data storage and management, resulting in poor recognition accuracy and adaptability.
By employing dynamic noise filtering and intelligent region segmentation, contour, color, and pixel density parameters are extracted through pixel-level dynamic threshold screening and multi-dimensional segmentation algorithms to construct a feature parameter set, which is then compared with a standard library. Combined with the parameter library update mechanism, adaptive feature matching is achieved.
It improves image recognition accuracy and robustness, enhances system adaptability and reliability, provides process traceability and data management capabilities, and supports system integration and expansion.
Smart Images
Figure CN122312420A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of image processing and computer vision technology, specifically artificial intelligence image analysis. Background Technology
[0002] In current applications such as industrial inspection, security monitoring, autonomous driving, and medical imaging, efficient and accurate image analysis is crucial for achieving automatic recognition and decision-making. Traditional image analysis methods typically rely on manually designed feature extraction operators (such as SIFT and HOG) combined with classifiers. The process generally includes image preprocessing, feature extraction, feature matching, and result output. However, these methods often exhibit insufficient robustness and poor adaptability when faced with real-world conditions such as varying lighting, noise interference, target deformation, and complex backgrounds. For example, in low-light or high-noise environments, traditional filtering methods easily lose detailed features in the image or fail to effectively distinguish noise from valid information, leading to a decrease in the accuracy of subsequent feature extraction.
[0003] Furthermore, existing region segmentation technologies often rely on single threshold segmentation or edge detection algorithms, which limits the utilization of image grayscale and texture information, easily leading to oversegmentation or undersegmentation and affecting the integrity of feature regions. In terms of feature parameter construction, traditional methods typically extract only single-dimensional features such as contour and color, lacking the fusion and weight optimization of multiple feature parameters. This results in insufficient feature representation capabilities, making them susceptible to interference from intra-class differences and inter-class similarities during matching and comparison, leading to increased false recognition rates.
[0004] On the other hand, most existing image analysis devices employ fixed-parameter processing flows, making it difficult to adapt to the dynamic changes required in different scenarios. For example, noise filtering thresholds are typically set to fixed values and cannot be adaptively adjusted based on the actual image quality; standard feature parameter libraries are also often statically configured, lacking the ability to iteratively optimize based on actual recognition results, and their matching accuracy tends to gradually decline after long-term use due to changes in the scene. Furthermore, most systems lack effective storage and management functions for analysis process and result data, hindering subsequent traceability, auditing, and model optimization.
[0005] Therefore, designing an image analysis method and device that can adapt to image quality, perform accurate region segmentation, fuse multi-dimensional features, and perform dynamic parameter optimization has become an important issue in improving the accuracy and robustness of image recognition in complex scenarios. Summary of the Invention
[0006] The purpose of this invention is to provide artificial intelligence image analysis, which accurately extracts and compares feature parameters through dynamic noise filtering and intelligent region segmentation, thereby achieving automatic and efficient recognition of target images, improving recognition accuracy and processing speed, and is suitable for complex scene analysis.
[0007] To achieve the above objectives, the present invention provides the following technical solution: artificial intelligence image analysis, comprising the following steps: acquiring the original image data of the target scene; performing pixel-level noise filtering on the original image data, retaining the feature information in the image while removing invalid interference pixels; performing region segmentation on the filtered image data, dividing at least two feature regions based on pixel grayscale difference and texture distribution patterns; extracting the contour parameters, color distribution parameters, and pixel density parameters of each feature region to construct a feature parameter set; comparing the feature parameter set with a preset standard parameter library, and outputting the feature recognition result of the target image through parameter matching degree calculation.
[0008] Furthermore, the pixel-level noise filtering process employs a dynamic threshold screening mechanism, which adjusts the noise filtering threshold in real time based on the average pixel grayscale value of the original image. The adjustment range of the threshold is ±15% of the average grayscale value of the original image.
[0009] Furthermore, in the region segmentation step, a multi-dimensional segmentation algorithm is adopted. First, preliminary segmentation is completed based on grayscale difference. Then, the regions after preliminary segmentation are merged or split by verifying the continuity of texture features to ensure the integrity of the feature regions.
[0010] Furthermore, in the process of constructing the feature parameter set, the contour parameters include the contour perimeter, contour area, and contour concavity / convexity; the color distribution parameters include the mean and variance of the three RGB colors within the region; and the pixel density parameters include the number of effective pixels per unit area and the uniformity of pixel distribution.
[0011] Furthermore, the parameter matching degree calculation adopts a weighted summation method, in which the contour parameter accounts for 40% of the weight, the color distribution parameter accounts for 35% of the weight, the pixel density parameter accounts for 25% of the weight, and the matching degree threshold is set to 85%. When the calculation result is not lower than the threshold, it is determined to be a successful match.
[0012] Further, it includes: a data acquisition module for acquiring raw image data of the target scene and converting the raw image data into a processable digital signal; a noise filtering module connected to the data acquisition module for performing pixel-level noise filtering on the image data in digital signal form and dynamically adjusting the filtering threshold; a region segmentation module connected to the noise filtering module for performing multi-dimensional region segmentation on the filtered image data and outputting feature region data; a parameter extraction module connected to the region segmentation module for extracting multiple types of parameters of the feature region and constructing a feature parameter set; and a comparison output module connected to the parameter extraction module for comparing the feature parameter set with a standard parameter library, calculating the matching degree, and outputting the recognition result.
[0013] Furthermore, the noise filtering module includes a threshold calculation unit and a pixel filtering unit. The threshold calculation unit is used to calculate a dynamic filtering threshold based on the grayscale mean of the original image, and the pixel filtering unit is used to remove invalid interference pixels and retain valid feature pixels based on the threshold.
[0014] Furthermore, the region segmentation module includes a preliminary segmentation unit and a verification and optimization unit. The preliminary segmentation unit completes preliminary image segmentation based on grayscale differences, and the verification and optimization unit optimizes and adjusts the segmented region through texture continuity analysis to eliminate segmentation deviations.
[0015] Furthermore, it also includes a parameter library update module, which is connected to the comparison output module. It is used to iteratively update the preset standard parameter library based on each comparison result to optimize the accuracy of parameter matching. The update frequency is once after every 100 image analyses are completed.
[0016] Furthermore, the comparison output module also includes a result feedback unit, which is used to feed back the feature recognition result to an external terminal in the form of a digital signal, and at the same time store the image data and feature parameters of this analysis to form a historical analysis database for easy subsequent tracing and querying.
[0017] This invention provides artificial intelligence image analysis, which has the following beneficial effects: 1. Significantly Improved Image Recognition Accuracy. By applying pixel-level dynamic threshold noise filtering to the original image, invalid interference introduced by the shooting environment and sensor noise can be effectively eliminated, resulting in cleaner image data while preserving the core image feature information. This method abandons the traditional fixed threshold approach and can adaptively adjust the filtering intensity according to the image's own grayscale characteristics, avoiding feature loss or noise residue that may be caused by a "one-size-fits-all" approach. Combined with subsequent region segmentation and multi-parameter extraction, the accuracy of the processed object is guaranteed from the underlying data level, laying a reliable data foundation for high-precision feature recognition and improving the anti-interference ability and overall confidence of the recognition results.
[0018] 2. Comprehensive and Robust Feature Extraction. This method employs a step-by-step, multi-dimensional feature analysis and extraction strategy. First, a multi-dimensional segmentation algorithm based on grayscale and texture intelligently identifies and divides feature regions with different attributes in the image, ensuring the integrity of the target region. Based on this, three key parameters—contour, color, and pixel density—are extracted simultaneously to form a comprehensive feature description set, achieving a quantitative representation of the target's shape, appearance, and internal structure. This multi-parameter system can more comprehensively characterize target features, effectively addressing feature fluctuations caused by complex situations such as shooting angle, lighting changes, and partial occlusion, enhancing the robustness and adaptability of the entire analysis process.
[0019] 3. Algorithm Adaptability and Evolvability. This scheme incorporates a dynamic adjustment and self-optimization mechanism. In the noise filtering stage, the threshold is dynamically calculated based on the image content, ensuring the adaptability of the processing strategy. More importantly, by adding a parameter library update module, the system can periodically iteratively optimize the standard parameter library based on historical comparison results, allowing the feature matching benchmark to continuously evolve with the accumulation of application scenarios and the enrichment of data samples. This self-learning and optimization capability enables the system to continuously approach a better recognition model, effectively responding to newly emerging target shapes or environmental changes, maintaining stable and improved analysis accuracy over the long term, and extending the effective lifespan of the system.
[0020] 4. Traceability and Auditability of Process Results. The key steps and outputs of the entire analysis process are well-documented. From the original image, filtered image, segmented region, to the extracted feature parameter set, and even the comparison calculation process with the standard library, a structured data chain is formed. The results feedback unit archives the complete recognition results, process data, and key images, building a historical analysis database. This is not merely simple result storage, but a record of the complete chain of analytical evidence, facilitating the tracing, reproduction, and auditing of each step of the analysis when recognition discrepancies arise or quality checks are required. This enhances the transparency and credibility of the system application, meeting the management requirements of high-reliability application scenarios.
[0021] 5. High system integration and flexible output formats. This solution modularizes the complete image analysis process, defining clear functional modules for data acquisition, processing, analysis, and output. The interfaces between modules are clearly defined, facilitating system integration and functional expansion. At the output level, it not only generates the final feature recognition results but also provides real-time feedback to external terminals in the form of digital signals, enabling rapid data exchange and linkage with other control systems or information platforms. Simultaneously, the system addresses the dual needs of real-time feedback and long-term archiving, ensuring efficient information transmission while building a knowledge accumulation database, providing stable and versatile data support for upper-level decision-making, big data analysis, or continuous model training. Attached Figure Description
[0022] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.
[0023] Figure 1 This is a flowchart of the image analysis method of the present invention; Figure 2 This is a flowchart of the pixel-level noise filtering process of the present invention; Figure 3 This is a flowchart of the image region segmentation process of the present invention; Figure 4 This is a flowchart of the feature parameter construction and matching process of the present invention; Figure 5 This is a flowchart illustrating the operation and iteration of the image analysis device of the present invention. Detailed Implementation
[0024] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses consistent with some aspects of this disclosure as detailed in the appended claims.
[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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.
[0026] How to use: Step 1: Data Acquisition and Input Activate the device's "data acquisition module." Point the device at the target scene to be analyzed; this module will automatically acquire raw image data and convert it into digital signals. Ensure the target is clear and complete in the image.
[0027] Step 2: Activate noise filtering The digital signal will be automatically transmitted to the "noise filtering module". This module uses a dynamic threshold filtering mechanism for pixel-level processing.
[0028] Core mechanism: The module first calculates the average pixel grayscale value of the original image, and then uses this average value as a benchmark to dynamically determine the noise filtering threshold within a range of ±15%.
[0029] Processing effect: Based on this real-time threshold, the module will accurately remove invalid interference pixels in the image, while retaining the effective pixels that constitute the target features to the greatest extent.
[0030] Step 3: Perform region segmentation The filtered image data enters the "region segmentation module." This module performs multi-dimensional segmentation: Preliminary segmentation: First, based on the gray-level differences between pixels, the image is initially divided into multiple blocks.
[0031] Verification and Optimization: Subsequently, the module analyzes the continuity of texture feature distribution within each block. Based on this analysis, necessary merging or splitting operations are performed on the blocks generated from the initial segmentation to ultimately divide at least two "feature regions" with clear boundaries and consistent internal features, ensuring the integrity of each feature region.
[0032] Step 4: Extract feature parameters The "parameter extraction module" will analyze each feature region obtained in the previous step and extract the following three types of parameters to jointly construct the "feature parameter set" of the target: Contour parameters: including contour perimeter, contour area, and contour convexity / concavity.
[0033] Color distribution parameters: calculate the mean and variance of the RGB (red, green, blue) components within the calculated area.
[0034] Pixel density parameter: counts the number of effective pixels per unit area and calculates the uniformity of pixel distribution.
[0035] Step 5: Compare and Output Results The "Comparison Output Module" compares the "Feature Parameter Set" generated in the previous step with the "Standard Parameter Library" pre-stored in the device.
[0036] Matching degree calculation: A weighted summation algorithm is used to calculate the matching degree. Among them, the contour parameter has a weight of 40%, the color distribution parameter has a weight of 35%, and the pixel density parameter has a weight of 25%.
[0037] Result determination: The system's preset matching success threshold is 85%. When the calculated result is not lower than this threshold, the module determines that the match is successful and outputs the final feature recognition result.
[0038] Results Feedback: The recognition results will be sent in digital signal form to the user-specified external terminal (such as a computer or monitor) via the "Result Feedback Unit". Simultaneously, the image data, feature parameters, and results of this analysis will be automatically stored, forming a historical record for easy tracing and retrieval later. Example Example 1
[0039] This embodiment provides an image analysis method, specifically applied to the visual sorting of parts of specific specifications on an industrial production line.
[0040] First, raw image data of the parts on the conveyor belt is acquired using a high-definition industrial camera. The method then performs pixel-level noise filtering on the acquired raw image data. This processing employs a dynamic threshold filtering mechanism; the system automatically calculates the average pixel grayscale value of the entire part image and, based on this average, calculates a specific threshold for this filtering operation in real time. According to this threshold, invalid interference pixels caused by environmental dust, light reflection, etc., are effectively removed from the image, while key feature information such as the edges and surface texture of the part itself is fully preserved.
[0041] Next, region segmentation is performed on the noise-filtered image data. The segmentation process consists of two stages: preliminary segmentation based on pixel grayscale differences, which roughly distinguishes the main parts, the background conveyor belt, and potential oil stain areas in the image; subsequently, the preliminary segmentation results are verified and optimized by analyzing whether the distribution of texture features in each region is continuous and uniform. For example, the main parts regions that were incorrectly segmented into multiple pieces due to reflection are re-merged, or the stuck parts and oil stain areas are accurately separated, ultimately dividing the main parts, qualified marking areas, and other well-defined feature regions.
[0042] Subsequently, the system extracts various parameters from each feature region to construct a feature parameter set. For the main body region of the part, its contour perimeter, area, and contour convexity reflecting the shape regularity are extracted as contour parameters; the RGB color distribution of its surface is analyzed, and the mean and variance are calculated as color distribution parameters; the number of effective pixels representing material texture per unit area and the uniformity of their distribution are counted as pixel density parameters. For the qualified marking area, the corresponding contour, color, and pixel density parameters are also extracted, together forming a complete feature parameter set for the part.
[0043] Finally, the set of feature parameters is compared with a preset standard parameter library (which stores the standard parameters of qualified parts). The system uses a weighted summation method to calculate the parameter matching degree, where contour parameters, color distribution parameters, and pixel density parameters are assigned different weights. When the calculated overall matching degree is not lower than the preset matching degree threshold, the system determines that the part is qualified and outputs the instruction "feature recognition result is qualified," triggering the subsequent robotic arm to sort it to the qualified product area. This method, through multi-step collaboration, achieves accurate and stable identification of part features in complex industrial environments. Example 2
[0044] This embodiment provides an image analysis method, specifically applied in the agricultural field for the automated identification of fruit maturity in orchards.
[0045] At the start of the operation, a camera mounted on the inspection equipment acquires raw image data of the fruit tree canopy. The method first performs pixel-level noise filtering. Considering the variable outdoor natural light environment, the system dynamically calculates the overall grayscale mean of the current image and adaptively adjusts the noise filtering threshold range accordingly. Using this dynamic threshold, interfering pixels introduced by sky highlights, leaf shadows, and distant background clutter can be effectively filtered out, while ensuring that the true color and texture information of the fruit and leaves are not destroyed and are clearly preserved.
[0046] Next, the system proceeds to the region segmentation step. It employs a multi-dimensional segmentation algorithm: first, based on differences in color and brightness (i.e., grayscale differences), it initially segments multiple regions from the image, such as suspected fruits, leaves, sky, and branches. Then, it verifies the continuity of texture features for each initially segmented region. For example, it merges the same fruit region that is broken due to occlusion but has continuous texture features (such as the distribution pattern of skin spots); or it separates leaf and insect-damaged fruit regions with significantly different texture features. After this optimization, several feature regions with accurate boundaries and uniform internal features are finally obtained, such as ripe fruit regions, unripe fruit regions, and healthy leaf regions.
[0047] Then, for the segmented "ripe fruit area," a detailed set of feature parameters is extracted. Contour parameters include the fruit's projected perimeter, approximate area, and the convexity / concavity reflecting the fruit's shape and fullness. Color distribution parameters focus on analyzing the RGB components of red, yellow, and other colors that characterize ripeness, calculating their mean and variance. Pixel density parameters quantify the gloss and texture of the fruit surface, achieved by calculating the effective number of pixels per unit area and their distribution uniformity. These parameters collectively construct a feature set describing the current ripeness state of the fruit.
[0048] Next, the set of feature parameters is compared with standard feature models for different maturity levels (such as "unripe," "color-changing," "ripe," and "overripe") in a standard parameter library. Through specific weighted matching degree calculations, the system calculates the matching degree between the current fruit features and each maturity model. When the matching degree calculation result with the "ripe" model reaches or exceeds a preset success threshold, the system outputs the feature recognition result as "ripe fruit" and can also record its location information, providing a basis for automated harvesting decisions. This method achieves intelligent, non-contact identification of fruit maturity in non-standard natural scenarios. Example 3
[0049] This embodiment provides an image analysis method, specifically applied to the identification of specific types of vehicles (such as hazardous chemical transport vehicles) in security monitoring.
[0050] The system acquires vehicle images from video streams in real time from surveillance cameras deployed at intersections as raw image data. To address image noise caused by complex nighttime lighting and weather conditions, the method initiates pixel-level noise filtering. Based on the average grayscale value of the current frame's vehicle image, the system determines a dynamic filtering threshold in real time. This threshold effectively suppresses invalid interfering pixels caused by low-light noise, headlight glare, and rain / snow lines, while systematically preserving key detail pixels such as the vehicle's outline, body markings, and license plate.
[0051] Next, region segmentation is performed on the filtered, sharpened image. Initial segmentation is achieved based on significant grayscale differences between the vehicle, road surface, and background buildings, separating the vehicle from the scene. Then, texture continuity is verified within the vehicle region: for example, cargo box areas with continuous regular stripes or reflective features are merged into a single feature region; license plate areas and hazardous materials sign areas with textures distinct from the vehicle body are clearly separated. After processing by a multi-dimensional segmentation algorithm, the image is divided into multiple feature regions, including the vehicle body, license plate, and special signs.
[0052] The construction of the feature parameter set will be carried out for different feature regions. For the main body area of the vehicle, its macroscopic contour parameters (such as aspect ratio and area) and color distribution parameters (such as the RGB mean of the main color) are extracted. For the hazardous materials sign area, its specific shape contour parameters (such as the perimeter and concavity of the rhombus), legally defined color RGB parameters and their variance, and pixel density and uniformity parameters of the internal pattern of the sign are precisely extracted. All these parameters are summarized into the feature parameter set of the current vehicle.
[0053] In the final identification stage, the set of feature parameters is compared with the feature templates of "hazardous materials transport vehicles" in the standard parameter library. The system calculates the comprehensive matching degree according to preset weight ratios (with different emphases on contour, color, and pixel density parameters). If the calculation result is not lower than the preset identification threshold, the comparison output module determines that the match is successful, outputs the feature identification result "hazardous materials transport vehicle identified", and can trigger an alert or record. This method improves the accuracy and reliability of automated identification of specific target vehicles in complex monitoring environments. Example 4
[0054] This embodiment provides an image analysis method, specifically applied in medical auxiliary diagnosis, to perform preliminary analysis on dermoscopy images to identify the characteristics of specific skin lesions.
[0055] First, raw image data of the patient's skin lesions is acquired using a digital dermoscopy device. Due to interference from skin surface contaminants such as dandruff, hair, and oil reflections, the method first performs pixel-level noise filtering. The system calculates the mean value based on the overall grayscale level of the dermoscopy image and dynamically generates a filtering threshold to accurately filter out interfering pixels from non-lesion structures, while ensuring that all important diagnostic features, such as pigment distribution, edge morphology, and vascular texture, in the lesion area are preserved without loss.
[0056] Next, region segmentation is performed on the processed image. Based on the differences in color and brightness (grayscale difference) between the lesion area and normal skin, the algorithm performs preliminary segmentation, outlining the approximate range of the suspected lesion. Subsequently, a more refined analysis of texture distribution patterns is conducted, and the preliminary segmentation results are optimized by verifying the continuity of textures. For example, sub-regions belonging to the same lesion but initially separated due to uneven color are merged; or the area where the lesion edge is separated from a single large hair is distinguished. Finally, the image is segmented into characteristic regions such as the lesion core area, the edge diffusion area, and the surrounding normal skin area.
[0057] The system then focuses on extracting a set of feature parameters from the core and peripheral diffusion areas of the lesion. Contour parameters include the perimeter of the lesion's irregular boundary, its total area, and the contour's convexity / concavity used to quantify peripheral infiltration. Color distribution parameters delve into the RGB mean and variance of different pigmentation regions within the lesion, which is particularly important for differential diagnosis. Pixel density parameters quantify subtle structural changes in the skin by calculating the number and uniformity of effective pixels reflecting texture per unit area. These multi-dimensional parameters constitute a digital feature set describing the skin lesion.
[0058] Subsequently, the set of feature parameters is compared with a pre-set standard parameter library. The standard parameter library stores feature parameter models for various typical skin lesions (such as common moles, melanoma, hemangiomas, etc.). The comparison output module calculates the feature matching degree between the current lesion features and each model using weighted averages. When the matching degree calculation result with a certain disease model (such as melanoma) reaches or exceeds a pre-set judgment threshold, the system outputs a suggestive feature recognition result, such as "Feature highly matches XX model," providing doctors with objective quantitative reference. This method achieves standardized extraction and auxiliary analysis of skin image features. Example 5
[0059] This embodiment provides an image analysis method, specifically applied to the authentication of specific seals or signatures in document digitization management.
[0060] First, the original image data of the document to be inspected is acquired using a high-resolution scanner. To address interference from yellowed paper, printing patterns, stains, creases, and other defects, the method performs pixel-level noise filtering. The system automatically calculates the overall grayscale mean of the document image and dynamically sets the filtering threshold. This operation effectively removes invalid interfering pixels such as paper background noise and irrelevant printed text, while preserving the red ink areas of the stamp or the ink pixels of the signature / signature clearly and completely.
[0061] Next, the algorithm proceeds to the region segmentation step. Based on the significant grayscale and color differences between the stamp (or handwriting) and the paper background, the algorithm performs initial segmentation, locating all possible suspicious stamp areas. Then, it uses texture distribution patterns for continuity verification and optimization. For stamps, it verifies the continuity and uniformity of the ink color distribution, reasonably connecting any breaks caused by improper stamping into a complete feature region. For handwriting, it analyzes the continuity of the ink texture, distinguishing it from accidental stains. Finally, it accurately segments the stamp or signature areas that need to be identified.
[0062] Next, refined feature parameters are extracted from the segmented target regions to construct a feature parameter set. For seals, contour parameters include the perimeter of the outer edge, the area of the internal pattern, and the concavity and convexity of each part; color distribution parameters precisely analyze the mean and variance of the RGB channels of the red ink; pixel density parameters quantify the distribution of ink density (effective pixel density) and the uniformity of stamping pressure (pixel distribution uniformity). For handwriting, the overall character contour parameters, ink depth (color distribution parameters), and details such as brushstrokes and dry brushstrokes are extracted (pixel density parameters).
[0063] Finally, the set of feature parameters of the target to be verified (such as seal A) is compared with the standard feature data of a genuine seal A that has been registered in the standard parameter library. The comparison output module uses a preset weight ratio to calculate the comprehensive matching degree between the two in terms of contour shape, color composition, micro-texture density, and other aspects. Only when the calculation result reaches or exceeds a strictly set matching degree success threshold is the system determined to be a successful match and outputs the conclusion "feature recognition result is true". Otherwise, a mismatch is indicated. This method provides an objective and quantitative image feature comparison means for document anti-counterfeiting.
[0064] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. An image analysis method, characterized in that, The process includes the following steps: acquiring the original image data of the target scene; performing pixel-level noise filtering on the original image data to retain the feature information in the image while removing invalid interference pixels; performing region segmentation on the filtered image data, dividing at least two feature regions based on pixel grayscale difference and texture distribution patterns; extracting the contour parameters, color distribution parameters, and pixel density parameters of each feature region to construct a feature parameter set; comparing the feature parameter set with a preset standard parameter library, and outputting the feature recognition result of the target image through parameter matching degree calculation.
2. The image analysis method according to claim 1, characterized in that, The pixel-level noise filtering process employs a dynamic threshold screening mechanism, which adjusts the noise filtering threshold in real time based on the average pixel grayscale value of the original image. The adjustment range of the threshold is ±15% of the average grayscale value of the original image.
3. The image analysis method according to claim 1, characterized in that, In the region segmentation step, a multi-dimensional segmentation algorithm is used. First, preliminary segmentation is completed based on grayscale difference. Then, the regions after preliminary segmentation are merged or split by checking the continuity of texture features to ensure the integrity of the feature regions.
4. The image analysis method according to claim 1, characterized in that, In the process of constructing the feature parameter set, the contour parameters include the contour perimeter, contour area and contour concavity / convexity, the color distribution parameters include the mean and variance of the three RGB colors in the region, and the pixel density parameters include the number of effective pixels per unit area and the uniformity of pixel distribution.
5. The image analysis method according to claim 1, characterized in that, The parameter matching degree calculation adopts a weighted summation method, in which the contour parameter accounts for 40% of the weight, the color distribution parameter accounts for 35% of the weight, the pixel density parameter accounts for 25% of the weight, and the matching degree threshold is set to 85%. When the calculation result is not lower than the threshold, it is determined to be a successful match.
6. An image analysis device, characterized in that, include: The data acquisition module is used to collect raw image data of the target scene and convert the raw image data into a processable digital signal; The noise filtering module, connected to the data acquisition module, is used to perform pixel-level noise filtering on image data in digital signal form and dynamically adjust the filtering threshold; the region segmentation module, connected to the noise filtering module, is used to perform multi-dimensional region segmentation on the filtered image data and output feature region data; the parameter extraction module, connected to the region segmentation module, is used to extract multiple types of parameters of the feature region and construct a feature parameter set. The comparison output module is connected to the parameter extraction module and is used to compare the feature parameter set with the standard parameter library, calculate the matching degree, and output the recognition result.
7. The image analysis apparatus according to claim 6, characterized in that, The noise filtering module includes a threshold calculation unit and a pixel filtering unit. The threshold calculation unit is used to calculate a dynamic filtering threshold based on the grayscale mean of the original image. The pixel filtering unit is used to remove invalid interference pixels and retain valid feature pixels based on the threshold.
8. The image analysis apparatus according to claim 6, characterized in that, The region segmentation module includes a preliminary segmentation unit and a verification and optimization unit. The preliminary segmentation unit completes preliminary image segmentation based on grayscale difference, and the verification and optimization unit optimizes and adjusts the segmented region through texture continuity analysis to eliminate segmentation deviation.
9. The image analysis device according to claim 6, characterized in that, It also includes a parameter library update module, which is connected to the comparison output module. It is used to iteratively update the preset standard parameter library based on each comparison result to optimize the accuracy of parameter matching. The update frequency is once after every 100 image analyses are completed.
10. The image analysis apparatus according to claim 6, characterized in that, The comparison output module also includes a result feedback unit, which is used to feed back the feature recognition results to an external terminal in the form of digital signals, and at the same time store the image data and feature parameters of this analysis to form a historical analysis database for easy subsequent tracing and querying.