A Structured Feature Parameter Extraction Method Based on Computer Vision

Abstract

This invention relates to the field of computer vision and feature extraction technology, specifically disclosing a structured feature parameter extraction method based on computer vision, comprising the following steps: S1: Projecting adaptive markers onto the surface of a target object using a tunable laser source, and acquiring images using a multi-band filtering module to generate an enhanced image; S2: Using a CNN branch to extract local texture features, using a Transformer branch to extract global spatial relationship features, and fusing local texture features and global spatial relationship features across scales using a feature pyramid to output a fused feature map; S3: Performing rasterization processing on the fused feature map and calculating the feature density, dynamically adjusting the feature point distribution based on the feature density to achieve uniform feature coverage; S4: Extracting candidate features through block cascading and accurately extracting and processing to output structured feature parameters, and combining model quantization to accelerate and improve the real-time performance of feature extraction; achieving accurate and efficient extraction of structured feature parameters.

Description

Technical Field

[0001] This invention relates to the field of computer vision and feature extraction technology, specifically to a method for extracting structured feature parameters based on computer vision. Background Technology

[0002] In the field of computer vision, feature extraction is the core foundation for downstream tasks such as target recognition, defect detection, and pose estimation. Existing feature extraction techniques suffer from several shortcomings: features are not prominent in weakly textured scenes; surfaces with weak textures, such as metal and glass, lack natural features, making it difficult for traditional vision methods to extract effective features, leading to high false negative rates; feature extraction is one-sided: a single network can only extract local textures, making it difficult to capture global spatial relationships, and features of small targets are easily lost; feature distribution is uneven: feature points tend to cluster in high-contrast areas, and there may be no effective features at image edges or in the center, affecting the stability of downstream tasks; real-time performance is poor: complex feature extraction models have high computational costs, making it difficult to meet real-time processing requirements on embedded devices (such as industrial robots); and there is a lack of quantitative evaluation and closed-loop optimization: feature quality cannot be objectively evaluated, and extraction strategies cannot be dynamically adjusted based on evaluation results, requiring manual intervention for optimization.

[0003] To address the aforementioned issues, this invention proposes a structured feature parameter extraction method that enhances the robustness, comprehensiveness, and practicality of feature extraction. Summary of the Invention

[0004] To overcome the aforementioned technical problems in the prior art, this invention provides a structured feature parameter extraction method based on computer vision. Through a closed-loop process of "image enhancement - feature fusion - distribution optimization - real-time output - evaluation optimization", the method achieves accurate and efficient extraction of structured feature parameters.

[0005] To achieve the above objectives, this invention provides a structured feature parameter extraction method based on computer vision, comprising the following steps: S1: Projecting adaptive markers onto the surface of a target object using a tunable laser source, and acquiring images using a multi-band filtering module to generate an enhanced image; S2: Extracting features from the enhanced image, constructing a CNN branch to extract local texture features, constructing a Transformer branch to extract global spatial relationship features, and fusing local texture features and global spatial relationship features across scales using a feature pyramid to output a fused feature map; S3: Performing rasterization processing on the fused feature map and calculating feature density, dynamically adjusting the feature point distribution based on the feature density to ensure uniform feature coverage; S4: Extracting candidate features through block cascading and accurately extracting and processing to output structured feature parameters, and combining model quantization to accelerate and improve the real-time performance of feature extraction.

[0006] Preferably, S1 specifically includes: the tunable laser source emitting a laser to the surface of the target object, projecting the adaptive mark onto the weakly textured surface of the target object, for generating a high-contrast mark on the weakly textured surface; the adaptive mark is a grid or dot matrix, and the multi-band filtering module includes at least four types of filters; the tunable laser source automatically adjusts the density and shape of the projected grid or dot matrix according to the surface texture of the target object; and the multi-band filtering module automatically switches filters according to ambient lighting conditions and the surface characteristics of the target object.

[0007] Preferably, in step S2, feature extraction is performed on the enhanced image. A CNN branch is used to extract local texture features, and a Transformer branch is used to extract global spatial relationship features. Specifically, the CNN branch uses a lightweight MobileNetV3 network structure to extract local texture features of the enhanced image, including scratches and solder joints. The Transformer branch uses a self-attention mechanism to extract global spatial relationship features of the enhanced image, including component layout. An information interaction module is provided between the CNN branch and the Transformer branch for feature information sharing and fusion between the two branches.

[0008] Preferably, in step S2, a fused feature map is generated by fusing local texture features and global spatial relationship features across scales using a feature pyramid. Specifically, this includes: fusing low-level feature maps of the CNN branch and high-level feature maps of the Transformer branch across scales according to the FPN structure of the feature pyramid to generate a fused feature map; and when fusing features, a weighted fusion algorithm is used to assign weights according to the contribution of features at different scales to the extraction of target features.

[0009] Preferably, S3 specifically includes: dividing the fused feature map into M×N grids: M,N≥5, and calculating the feature density index using formula (1);

[0010] Formula (1);

[0011] in, The number of feature points, , For location variance;

[0012] During the rasterization process, for rasters with a feature density index lower than a preset threshold, a feature enhancement algorithm is used to supplement feature points; a dynamic distance constraint operation is performed to force the spacing between feature points to be greater than 20 pixels within the rasters selected based on the feature density index.

[0013] Preferably, S4 specifically includes: a first-level processing: reducing the resolution of the fused feature map to 1 / 4 of the original resolution to obtain a low-resolution image, and using the FAST algorithm to quickly extract candidate features on the low-resolution image; a second-level processing: performing fine feature extraction on the original resolution image for the region where the candidate features extracted in the first-level processing are located, and outputting the structured feature parameters; the structured feature parameters include geometric features, texture features and semantic features.

[0014] Preferably, the method further includes: performing quality evaluation on the structured feature parameters based on evaluation indicators and generating a quality assessment report; the evaluation indicators include illumination fluctuation tolerance, feature repetition rate, and grid coverage and efficiency; the quality assessment report includes four indicator data: (1) robustness indicator: illumination fluctuation tolerance > 500 lux; (2) accuracy indicator: feature repetition rate > 95%, small target mAP@0.5 > 75%; (3) distribution indicator: grid coverage > 90%; (4) efficiency indicator: single frame processing latency < 40 ms.

[0015] Preferably, in step S3, when the grid coverage is detected to be <90%, the density index threshold is automatically reduced; when the feature repetition rate is <95%, the shallow feature weights of the CNN branch are increased; when the illumination fluctuation tolerance is <500 lux, an anti-interference filter is switched to be used; when the small target mAP@0.5 is <75%, the weights of the lower layers of the feature pyramid are increased.

[0016] The present invention has at least the following technical effects through the technical solution provided by the present invention:

[0017] Enhanced robustness: Active laser marking addresses the lack of features on weakly textured surfaces; multi-band filtering adapts to varying illumination, improving the tolerance for illumination fluctuations and increasing the feature detection rate in weakly textured scenes; Comprehensive features: Dual-branch network fuses local and global features, and cross-scale fusion avoids the loss of small target features, significantly improving the accuracy of small target detection; Uniform distribution: Rasterization optimization improves feature coverage, thus avoiding local missed detections and effectively reducing the error rate of downstream tasks; High real-time performance: Block-based cascaded extraction and model quantization acceleration effectively reduce single-frame processing latency, meeting the real-time detection needs of industrial production lines; Adaptive optimization: Automatic tuning is achieved through quantization evaluation and closed-loop adjustment, requiring no manual intervention and significantly improving feature quality stability. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings:

[0019] Figure 1 This is a flowchart of a structured feature parameter extraction method based on computer vision provided by an embodiment of the present invention;

[0020] Figure 2 This is a diagram of the CNN-Transformer feature fusion architecture provided in this embodiment of the invention;

[0021] Figure 3 This is a flowchart of the rasterization optimization process provided in an embodiment of the present invention;

[0022] Figure 4 This is a structural diagram of a computer vision-based structured feature parameter extraction system provided in a real-time example of the present invention. Detailed Implementation

[0023] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the scope of the present invention.

[0024] In this invention, the terms "system" and "network" are used interchangeably. "Multiple" refers to two or more; therefore, in this invention, "multiple" can also be understood as "at least two." "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / ", unless otherwise specified, generally indicates that the preceding and following related objects have an "or" relationship. Furthermore, it should be understood that in the description of this invention, terms such as "first" and "second" are used only for descriptive purposes and should not be construed as indicating or implying relative importance or order.

[0025] This invention provides a method for extracting structured feature parameters based on computer vision, such as... Figure 1 As shown, it includes the following steps:

[0026] S1: An adaptive marker is projected onto the surface of the target object using a tunable laser source, and an enhanced image is generated by combining the image with a multi-band filtering module.

[0027] S2: Perform feature extraction on the enhanced image, construct a CNN branch to extract local texture features, construct a Transformer branch to extract global spatial relationship features, and fuse local texture features and global spatial relationship features across scales through a feature pyramid to output a fused feature map;

[0028] S3: Perform rasterization processing on the fused feature map and calculate the feature density. Based on the feature density, dynamically adjust the distribution of feature points to make the feature coverage uniform.

[0029] S4: Extract candidate features through block cascading and accurately extract and process the output structured feature parameters, and combine model quantization to accelerate and improve the real-time performance of feature extraction.

[0030] In this embodiment of the invention, an enhanced image is generated through S1. Image features are enhanced through active labeling and multi-band adaptation, providing high-quality input for subsequent extraction. First, a tunable laser source emits an infrared laser to project adaptive markings onto the target surface. The wavelength of the infrared laser is 850nm. The adaptive markings are projected onto the weakly textured surface of the target object to generate high-contrast markings. For weakly textured surfaces such as metal and glass, the laser markings can generate high-contrast feature points, similar to drawing marks on a smooth blackboard. The adaptive markings can be a grid or a dot matrix. The tunable laser source can automatically adjust the density and shape of the projected grid or dot matrix according to the surface texture of the target object. For example, a sparse grid is used for rough surfaces, a dense dot matrix for smooth surfaces, and an adaptive dot matrix for irregular targets. Then, an industrial camera collects images through a multi-band filtering module. The multi-band filtering module includes at least four band filters, such as 450nm blue light, 550nm green light, 650nm red light, and 850nm... The near-infrared, multi-band filtering module can automatically switch filters according to ambient light (e.g., using a narrow-band filter to reduce glare in strong light and a wide-band filter to enhance brightness in weak light) and target characteristics (e.g., using a near-infrared filter to highlight edges on metal and a green light filter to highlight texture on plastic), ultimately generating an enhanced image that adapts to the environment.

[0031] In embodiments of the present invention, such as Figure 2As shown, features are extracted and fused using S2 to address the issue of incomplete feature extraction. A strategy combining dual-branch extraction and cross-scale fusion is employed to simultaneously capture local details and global relationships, avoiding feature omissions. Specifically, dual-branch feature extraction is performed on the enhanced image, with two branches: a CNN branch and a Transformer branch. The CNN branch is used to extract local texture features, specifically employing a lightweight MobileNetV3 network to focus on extracting local texture features in the enhanced image, such as scratches, weld points, and holes on the surface of parts. Its advantage is its strong sensitivity to local textures, making it suitable for capturing minute features. The Transformer branch is used to extract global spatial relationship features, specifically capturing global spatial relationship features of the enhanced image through a self-attention mechanism, such as the position of parts in the image and the layout relationship of multiple parts. Its advantage is its ability to model long-distance dependencies, avoiding the situation of seeing only the local and not the whole. The CNN branch and the Transformer branch communicate through an information interaction module. Time-sharing features are used. For example, the CNN passes the local scratch location to the Transformer, and the Transformer feeds back the global layout to the CNN, improving feature correlation. That is, the shallow features of the CNN (1 / 8 scale) are reduced in dimensionality by 1×1 convolution and then input into the key-value matrix of the Transformer. The high-level features of the Transformer are fed back into the residual connection layer of the CNN branch. Local texture features and global spatial relationship features are fused across scales through the feature pyramid to output a fused feature map. Specifically, the two types of features are fused according to the FPN structure of the feature pyramid. The low-level high-resolution features (including local texture features) output by the CNN and the high-level semantic features (including global spatial relationship features) output by the Transformer are fused and concatenated according to scale, and a weighted fusion algorithm is used (such as adding low-level feature weights to small target areas and adding high-level feature weights to global layout areas). That is, the fusion is performed using a fusion formula, which is:

[0032] F_fused=α×Conv1x1(F_cnn)+β×UpSample(F_trans),

[0033] Wherein, F_fused is the final fused feature map, Conv1x1(F_cnn) is the feature map after performing a 1×1 convolution (Conv1x1) operation on the feature map output by the CNN, and UpSample(F_trans) is the feature map after performing an upsampling (UpSample) operation on the feature map output by the Transformer. The operation weights α and β are generated by learnable parameters, with initial values ​​of α=0.6 (emphasizing texture) and β=0.4 (emphasizing spatial relationships). The final output is a fused feature map covering "local-global" and "detail-semantic" aspects.

[0034] In embodiments of the present invention, such as Figure 3 As shown, the feature distribution is optimized by S3, and the features are uniformly covered by the image through a combination of rasterization evaluation and dynamic adjustment to avoid local missed detections; the fused feature map is divided into M×N grids, where M and N≥5, for example, divided into 5×5 grids, and the feature density index of each grid is calculated by formula (1).

[0035] Formula (1);

[0036] in, The number of feature points, , For the location variance, the lower the density index calculated by formula (1), the sparser the features of the region. Furthermore, in the rasterization process, for rasters with a density index lower than a preset threshold (e.g., edge regions), a feature enhancement algorithm (e.g., local contrast enhancement) is used to supplement feature points. For example, if the preset threshold is 0.3, if the feature density index of raster A is ≥0.3, the original features of raster A are retained. If the feature density index of raster B is <0.3, feature enhancement is triggered. The execution steps of the feature enhancement algorithm are: first, edge enhancement, i.e., applying the Sobel operator to enhance the edge response within the raster; then, corner point supplementation, i.e., using Harris corner point detection to add new feature points; finally, spacing constraint, forcing the distance between the new points and the existing points to be >20 pixels, thereby achieving the visualization effect of adding 6 feature points to raster B; and then, performing dynamic distance constraint operation, specifically for rasters with clustered feature points, forcing the distance between feature points to be greater than 20 pixels. Pixels refer to the feature points that traverse all grid cells. If the distance between two points is ≤20 pixels, the one with the lower response value is removed. For example, if the original distance between two points is 15 pixels, after removing the last point, the distance is >20 pixels to avoid overlap. Ultimately, the feature is evenly covered in all grid cells, and an optimized feature map is output. In the optimized feature map, the feature density index of all grid cells is ≥ a preset threshold, the distance between any feature points is >20 pixels, and the grid coverage is >90%.

[0037] In this embodiment of the invention, structured feature parameters are output through S4. Efficient output is achieved through a combination of block extraction and quantization acceleration, ensuring that the features can be directly used for downstream tasks. The block-based cascaded extraction in S4 includes a first-level processing and a second-level processing. The first-level processing specifically involves reducing the resolution of the fused feature map to 1 / 4 of the original resolution to obtain a low-resolution image. Then, the FAST algorithm is used to quickly filter candidate features on the low-resolution image, such as excluding invalid features in the background region to reduce computational load. The second-level processing specifically involves accurately extracting structured feature parameters from the original resolution image for the regions containing candidate features obtained through the first-level processing. These structured feature parameters include geometric features (such as target size and position coordinates), texture features (such as scratch depth and surface roughness), and semantic features (such as part type and layout relationships). Lightweight acceleration can also be used, specifically employing INT8 quantization (converting model parameters from 32-bit floating-point to 8-bit integer) combined with TensorRT optimization to reduce model computation and ensure real-time performance on embedded devices, for example, single-frame processing latency <40ms.

[0038] In this embodiment of the invention, a combination of multi-dimensional evaluation and dynamic adjustment is also included to continuously optimize feature quality. Specifically, the quality evaluation of structured feature parameters is performed based on evaluation indicators, and a quality assessment report is generated. The evaluation indicators include illumination fluctuation tolerance, feature repetition rate, grid coverage and efficiency. The quality assessment report includes four indicator data, specifically including (1) robustness indicator: illumination fluctuation tolerance > 500 lux, that is, the feature is stable within 500 lux of illumination change; (2) accuracy indicator: feature repetition rate > 95%, that is, the feature matching success rate of the same target at different angles, and small target mAP@0.5 > 75%, that is, the small target detection accuracy reaches more than 75%; distribution indicator: grid coverage is greater than 90%, that is, more than 90% of the grid has effective features; efficiency indicator: single frame processing latency is less than 40ms; after receiving the quality assessment report, a closed-loop optimization strategy is also executed. Specifically, if the grid coverage is < 90%, the feature density index threshold is automatically reduced and the feature supplementation of sparse areas is increased; if the feature repetition rate is < 95%, CNN is added. Branch shallow layer feature weights (enhancing texture stability); if the illumination fluctuation tolerance is <500 lux, switch to an anti-interference filter (such as an 850nm near-infrared filter); if the mAP@0.5 of small targets is <75%, increase the weights of the lower layers of the feature pyramid (enhancing detailed features).

[0039] In one implementation, the following detailed description is provided in the context of an industrial PCB board inspection scenario. The requirement is to extract structured feature parameters from solder joints (local texture), component layout (global relationships), and microvias (small targets) on the PCB board for detecting solder joint defects and component misassembly. Real-time processing (single frame latency <40ms) is required in a pipeline (speed 3m / s). First, an enhanced image is generated using S1. An adaptive dot matrix is ​​projected onto the PCB board (including areas with weak texture copper foil) using a tunable laser source (wavelength 850nm). Specifically, a dense dot matrix (0.5mm spacing) is projected onto the copper foil area (weak texture), and a sparse dot matrix (2mm spacing) is projected onto the component area (strong texture). A multi-band filtering module includes wavelengths of 450nm, 550nm, 650nm, and 850nm. A filter is used: an 850nm filter is switched in strong light (anti-glare), and a 550nm filter is switched in cloudy light (brightness enhancement) to generate an enhanced PCB board image. Then, S2 is used to extract and fuse features. The CNN branch (MobileNetV3) extracts local features of solder joint texture (e.g., whether the solder joint is rounded) and micro-holes (diameter < 0.3mm); the Transformer branch extracts global features of component layout (e.g., the relative positions of resistors and capacitors). Feature pyramid fusion is used, specifically adding low-level feature weights (weight 0.8) to the micro-hole region and high-level feature weights (weight 0.7) to the component layout region, outputting a fused feature map. Next, S3 is used to optimize the feature distribution, dividing the fused feature map into a 5×5 grid (25 regions in total), calculating the feature density index ρi for each grid; feature points are added to edge grids (e.g., grids with ρi below the threshold of 0.3), forcing the spacing between feature points within all grids to be greater than 20. To avoid clustering, the final feature coverage reached 92%. Then, S4 was used to output structured feature parameters, which were extracted in a block-by-block cascade manner: first, candidate features (solder joints and component edges) were screened on the 1 / 4 resolution image using the FAST algorithm, and then precisely extracted on the original resolution image: geometric features: solder joint diameter (accuracy ±0.02mm), component coordinates (accuracy ±0.1mm); texture features: solder joint roughness (level 1-5), number of microvias; semantic features: component type (resistor / capacitor), layout compliance; model quantization: INT8 quantization and TensorRT optimization were used, with a processing latency of 35ms on the Jetson Xavier device (meeting the 40ms requirement); finally, the structured feature parameters were evaluated and optimized, with evaluation indicators: illumination fluctuation tolerance 550 lux (>500 lux), feature repetition rate 96% (>95%), grid coverage 92% (>90%), and latency 35ms (<40ms); since all indicators met the standards, no optimization was triggered, and the structured feature parameters were directly output to the PCB inspection system.

[0040] The present invention provides a structured feature parameter extraction method based on computer vision, which has the following beneficial effects:

[0041] 1. Strong robustness: It solves the problem of insufficient surface features in weak texture by active laser marking, multi-band filtering adapts to changes in illumination, the illumination fluctuation tolerance is increased to more than 500 lux, and the feature detection rate of weak texture scene is increased by more than 90%.

[0042] 2. Comprehensive features: The dual-branch network fuses local and global features, and cross-scale fusion avoids the loss of small target features, achieving a small target detection accuracy (mAP@0.5) of over 75%;

[0043] 3. Uniform distribution: Rasterization optimization achieves feature coverage of over 90%, avoiding local missed detections and reducing downstream task errors by 40%;

[0044] 4. High real-time performance: Block-based cascaded extraction and model quantization acceleration, single-frame processing latency <40ms, meeting the real-time detection needs of industrial production lines;

[0045] 5. Adaptive optimization: Quantitative evaluation and closed-loop adjustment enable automatic optimization without manual intervention, improving feature quality stability by 60%.

[0046] Furthermore, such as Figure 4 As shown, the present invention also provides a structured feature parameter extraction system based on computer vision, comprising:

[0047] The image enhancement module is configured to project adaptive marks onto the surface of a target object using a tunable laser source, and to acquire images using a multi-band filtering module to generate an enhanced image. The tunable laser source can adjust the density and shape of the marks according to the surface texture of the target object, and the multi-band filtering module can switch filters according to ambient light and target characteristics.

[0048] The feature fusion module is configured to perform feature extraction on the enhanced image, including extracting local texture features using a CNN branch (MobileNetV3 network) and extracting global spatial relationship features using a Transformer branch, and fusing the two types of features across scales through a feature pyramid to output a fused feature map. An information interaction module is provided between the CNN branch and the Transformer branch.

[0049] The distribution optimization module is configured to perform rasterization processing on the fused feature map (dividing it into M×N grids, where M, N≥5), calculate the feature density index, and dynamically adjust the feature point distribution based on the index, including supplementing feature points to low-density grids and forcing the feature point spacing to be greater than 20 pixels.

[0050] The feature output module is configured to extract candidate features through block concatenation and refine them to output structured feature parameters (including geometric features, texture features, and semantic features), and combine model quantization to accelerate and improve real-time performance.

[0051] The evaluation and optimization module is configured to perform quality evaluation on structured feature parameters based on evaluation indicators such as illumination fluctuation tolerance, feature repetition rate, raster coverage and efficiency, generate a quality assessment report, and automatically adjust the parameters of each module according to the report to optimize the feature extraction effect.

[0052] Furthermore, embodiments of the present invention also provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the present invention.

[0053] The optional embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the embodiments of the present invention are not limited to the specific details in the above embodiments. Within the scope of the technical concept of the embodiments of the present invention, various simple modifications can be made to the technical solutions of the embodiments of the present invention, and these simple modifications all fall within the protection scope of the embodiments of the present invention.

[0054] It should also be noted that the various specific technical features described in the above embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, the embodiments of the present invention will not describe the various possible combinations separately.

[0055] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a microcontroller, chip, or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0056] Furthermore, various different implementations of the present invention can be combined arbitrarily, as long as they do not violate the spirit of the present invention, they should also be regarded as the content disclosed in the present invention.

Claims

1. A method for extracting structured feature parameters based on computer vision, characterized in that, Includes the following steps: S1: An adaptive marker is projected onto the surface of the target object using a tunable laser source, and an enhanced image is generated by combining the image with a multi-band filtering module. S2: Perform feature extraction on the enhanced image, construct a CNN branch to extract local texture features, construct a Transformer branch to extract global spatial relationship features, and fuse local texture features and global spatial relationship features across scales through a feature pyramid to output a fused feature map; S3: Perform rasterization processing on the fused feature map and calculate the feature density. Based on the feature density, dynamically adjust the distribution of feature points to make the feature coverage uniform. S4: Extract candidate features through block cascading and accurately extract and process the output structured feature parameters. Combine this with model quantization to accelerate and improve the real-time performance of feature extraction. Specifically, this includes: First-level processing: The resolution of the fused feature map is reduced to 1 / 4 of the original resolution to obtain a low-resolution image, and candidate features are quickly extracted from the low-resolution image using the FAST algorithm. Second-level processing: For the regions where the candidate features extracted in the first-level processing are located, fine feature extraction is performed on the original resolution image, and the structured feature parameters are output. The structured feature parameters include geometric features, texture features, and semantic features.

2. The method for extracting structured feature parameters based on computer vision according to claim 1, characterized in that, S1 specifically includes: The tunable laser source emits a laser beam onto the surface of the target object and projects the adaptive mark onto the weak textured surface of the target object to generate a high-contrast mark on the weak textured surface. The adaptive marking is a grid or a dot matrix, and the multi-band filtering module contains at least four types of band filters; The tunable laser source automatically adjusts the density and shape of the projected grid or dot matrix according to the surface texture of the target object; The multi-band filtering module automatically switches filters based on ambient lighting conditions and the surface characteristics of the target object.

3. The method for extracting structured feature parameters based on computer vision according to claim 1, characterized in that, In step S2, feature extraction is performed on the enhanced image. This involves using a CNN branch to extract local texture features and a Transformer branch to extract global spatial relationship features. Specifically, this includes: The CNN branch is constructed using a lightweight MobileNetV3 network structure to extract local texture features from the enhanced image, including scratches and solder joints. The Transformer branch is constructed based on a self-attention mechanism to extract global spatial relationship features of the enhanced image, and the global spatial relationship features include part layout; An information interaction module is provided between the CNN branch and the Transformer branch. The information interaction module is used for feature information sharing and fusion between the two branches.

4. The method for extracting structured feature parameters based on computer vision according to claim 3, characterized in that, In step S2, a fused feature map is generated by fusing local texture features and global spatial relationship features across scales using a feature pyramid. Specifically, this includes: Based on the FPN structure of the feature pyramid, the low-level feature maps of the CNN branch and the high-level feature maps of the Transformer branch are spliced ​​and fused across scales to generate a fused feature map. When fusing features, a weighted fusion algorithm is used to assign weights based on the contribution of features at different scales to the extraction of target features.

5. The method for extracting structured feature parameters based on computer vision according to claim 1, characterized in that, S3 specifically includes: The fused feature map is divided into M×N grids (M, N≥5), and the feature density index is calculated using formula (1); Formula (1); Among them, The number of feature points, , For location variance; During the rasterization process, for rasters with a feature density index lower than a preset threshold, a feature enhancement algorithm is used to supplement feature points. Perform a dynamic distance constraint operation to force the spacing between feature points to be greater than 20 pixels within a grid selected based on the feature density index.

6. The method for extracting structured feature parameters based on computer vision according to claim 1, characterized in that, The method further includes: The structured feature parameters are evaluated based on the evaluation indicators, and a quality assessment report is generated. The evaluation metrics include illumination fluctuation tolerance, feature repetition rate, raster coverage, and efficiency. The quality assessment report includes four key data points: (1) Robustness index: Illumination fluctuation tolerance > 500 lux; (2) Accuracy indicators: Feature repetition rate > 95%, small target mAP@0.5 > 75%; (3) Distribution index: Raster coverage > 90%; (4) Efficiency index: Single frame processing latency < 40ms.

7. The method for extracting structured feature parameters based on computer vision according to claim 6, characterized in that: In S3, when the grid coverage is detected to be <90%, the density index threshold is automatically reduced; When the feature repetition rate is less than 95%, increase the weight of shallow features in the CNN branch; When the illumination fluctuation tolerance is <500 lux, switch to using an anti-interference filter; When the small target mAP@0.5 < 75%, Increase the weight of the lower layers of the feature pyramid.

Images