Light field surface defect identification system and method based on multi-angle scattering

Abstract

The invention discloses a light field surface defect identification system and method based on multi-angle scattering, and relates to the field of optical image analysis. The system comprises a multi-angle scattered light field acquisition module and a defect intelligent identification module. The acquisition module obtains a hemispherical space scattered light field image of an optical element to be detected through one-time exposure of a collimation laser light source, a hemispherical dome and the like. The recognition module integrates a data preprocessing unit and an SFD-YOLO network, and on the basis of YOLOv11, the network replaces CBS in a shallow layer with a PSiam attention module to strengthen local significant feature perception, and replaces C3K2 in a deep layer with an OfficitMambaCSP module to strengthen global semantic modeling. The method comprises the steps of image acquisition, data set generation preprocessing, network training and defect identification. The problems that in the prior art, single-view-angle information is limited, and scattering data is insufficient in utilization are solved, Top1ACC reaches 95.6%, the parameter quantity is only 1.25 M, and the method is suitable for high-precision real-time detection of optical element defects.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of optical image analysis, in particular to a light field surface defect identification system and method based on multi-angle scattering. BACKGROUND

[0002] Optical elements are core components in the fields of medical imaging, industrial manufacturing, national defense research, consumer electronics, etc., and their surface quality (such as surface roughness, surface accuracy, surface defects) directly determines the system performance and reliability. Among them, although microscopic defects such as scratches and pits are small in size, they can damage the integrity of the optical surface, cause light scattering, diffraction or abnormal absorption, and significantly degrade the imaging quality, machining precision or sensing performance. However, such defects have complex morphology, random distribution, and are easily confused with background texture, making detection extremely difficult.

[0003] The existing optical surface defect detection technology has the following shortcomings: optical microscopic imaging (such as dark-field confocal microscopy) has high precision but low efficiency, and relies on single-view detection, making it difficult to obtain the full-angle optical response of the defect; machine vision detection (such as batch detection system based on YOLO) is efficient, but the mainstream method relies on single-view planar image input, which cannot capture multi-angle scattering features in the hemispherical space, limiting the defect classification and size inversion capability. Bidirectional reflectance distribution function (BRDF) measurement technology (such as ellipsoidal reflection structure, semi-parabolic mirror system) has realized the synchronous acquisition of hemispherical space scattering light field, but existing research only focuses on the measurement and modeling of BRDF data, and does not solve the core problem of "inverting defect type / size based on scattering light field information", and fails to combine measurement data with intelligent recognition.

[0004] The existing defect recognition network (such as YOLOv11-cls) has insufficient feature extraction capability for scattering light field images — scattering images have the problems of insignificant local features (such as brightness concentration areas) and weak global semantics (such as light intensity distribution rules), which leads to low recognition accuracy of similar size defects, and makes it difficult to balance "recognition accuracy" and "model lightweight" (such as parameter quantity, computational overhead), and cannot meet the real-time detection needs of industrial scenarios.

[0005] Therefore, it is urgent to design an intelligent recognition scheme that can simultaneously acquire complete multi-angle scattering information and adapt to scattering image features, in order to solve the problems of "low acquisition efficiency, insufficient recognition accuracy, poor engineering applicability" in existing technology. SUMMARY

[0006] The present application relates to the technical field of optical image analysis, in particular to a light field surface defect identification system and method based on multi-angle scattering.

[0007] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0008] A multi-angle scattering-based light field surface defect recognition system, comprising an illumination system, a sample stage, a hemispherical dome, an aspherical mirror, a relay optical imaging system and a detector, a hardware unit and a software unit;

[0009] The illumination system is a collimated laser light source for outputting monochromatic laser light and incidenting to the surface of the optical element to be measured at a preset zenith angle, and the incident direction is perpendicular to the length direction of the defect.

[0010] The sample stage is arranged at the geometric center of the hemispherical dome for fixing the optical element to be measured.

[0011] The inner wall of the hemispherical dome is coated with a diffuse reflective coating for reflecting the scattered hemispherical space light from the surface of the optical element to be measured.

[0012] The aspherical mirror is located outside the hemispherical dome for reflecting the scattered light reflected by the dome to the relay optical imaging system.

[0013] The relay optical imaging system comprises a lens group and a diaphragm for focusing and imaging the scattered light to the target surface of the detector.

[0014] The detector is a CCD or CMOS sensor for converting the scattered light radiance distribution into a digital scattered light field image.

[0015] The hardware unit includes a Linux operating system computer, an NVIDIA GPU and an Intel CPU, supporting a PyTorch deep learning framework.

[0016] The software unit includes a data preprocessing module and an SFD-YOLO network; the data preprocessing module is used for normalizing and size standardizing the scattered light field image; the SFD-YOLO network is used for extracting scattered image features and outputting defect categories.

[0017] Specifically, the SFD-YOLO network takes YOLOv11-cls as the basic architecture, including an input layer, a backbone feature extraction module, a PSimam attention module, an Efficient_Mamba_CSP module, a C2PSA module and a classification head.

[0018] The backbone feature extraction module retains the first two layers of CBS convolution blocks of YOLOv11-cls, and the 2 CBS modules in the shallow back section are replaced by the PSimam attention module.

[0019] The PSimam attention module includes two parallel multi-scale convolution branches and a SimAM non-parametric attention unit.

[0020] The two C3K2 modules in the backbone feature extraction module are replaced by Efficient_Mamba_CSP modules; the Efficient_Mamba_CSP modules are based on the CSPNet structure, and the main branch contains deep separable convolution and HSM-SSD mechanism, and the residual branch directly transmits original features, for enhancing global semantic modeling.

[0021] A light field surface defect identification method based on multi-angle scattering, comprising the following steps:

[0022] S1: Multi-angle scattering light field image acquisition: adjust the zenith angle of the illumination system, turn on the detector for one-time exposure, and obtain the hemispherical space scattering light field image of the optical element to be measured;

[0023] S2: Data set generation and preprocessing: parameterized construction of scratch defect model, generating 1180 groups of samples; normalizing the samples, and scaling to 640*640 pixels through bicubic interpolation;

[0024] S3: SFD-YOLO network training: configure training parameters, perform end-to-end training based on the data set of step S2 until the model converges;

[0025] S4: Defect identification: input the preprocessed image of the image to be detected into the trained SFD-YOLO network, and output the defect category (scratch width level).

[0026] Specifically, the data set generation in step S2 includes: for 0.1~0.9mm width scratch, generating 100 groups of center position samples + 10 groups of X-axis positive / negative offset samples; for 1.0mm width scratch, generating 100 groups of center position samples; the number of sample light tracing is 2~3 million, and the energy error peak value is <5%.

[0027] Specifically, in step S3, the Efficient_Mamba_block extracts local features through deep separable convolution, maps the feature sequence to the compressed state space through the HSM-SSD mechanism, captures long-distance dependence, and the computational complexity is linear.

[0028] Beneficial effects:

[0029] The multi-angle scattering light field surface defect identification system shown in the present application can obtain the scattering light field information of the hemispherical space through one-time exposure, without the need for visual angle scanning and splicing, and the acquisition efficiency is improved by 3~5 times compared with the traditional BRDF measurement system, and the inconsistency problem caused by multiple acquisitions is avoided, providing complete data support for defect identification.

[0030] The SFD-YOLO network enhances local salient feature perception through the PSIMam module and improves global semantic modeling through the Efficient_Mamba_CSP module. It achieves a Top_1 ACC of 95.6% on a 10-class scratch defect dataset, which is 3.1% higher than the original YOLOv11-cls. In particular, it achieves 100% accuracy in recognizing fine scratches of 0.1~0.4mm, solving the problem of easy confusion between defects of similar size.

[0031] SFD-YOLO has only 1.25M parameters, which is 18.8% less than YOLOv11-cls. Its computational overhead (Flops) is 4.0G, and it can achieve real-time inference on conventional GPUs, meeting the batch detection needs of industrial scenarios. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of the multi-angle scattered light field measurement system shown in this invention;

[0033] Figure 2 This is a schematic diagram of the SFD-YOLO classification network architecture shown in this invention;

[0034] Figure 3 This is a schematic diagram of the PSimam architecture shown in this invention;

[0035] Figure 4 This is a schematic diagram of the Efficient_Mamba_block architecture shown in this invention;

[0036] Figure 5 This is a schematic diagram of the Efficient_Mamba_CSP architecture shown in this invention;

[0037] Figure 6 This is a schematic diagram of the scratch defect model shown in this invention;

[0038] Figure 6 a is an isometric view of the scratch defect model shown in this invention;

[0039] Figure 6 b is a schematic diagram of the longitudinal cross-sectional distribution of scratches in the scratch defect model shown in this invention;

[0040] Figure 7 This is a schematic diagram showing the comparison of the scattered light field image before and after processing according to the present invention;

[0041] Figure 8 This is a graph showing the variation of the Top_1 ACC of SFD-YOLO as a function of training rounds, as illustrated in this invention.

[0042] Figure 9 This is a diagram showing the scattering image classification prediction results of the present invention. Detailed Implementation

[0043] The technical solution 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, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. In the description of the present invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention. In addition, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance. In the description of the present invention, it should be noted that unless otherwise expressly specified and limited, the terms "installed," "connected," and "linked" should be interpreted broadly. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be a connection within two components. Those skilled in the art will understand the specific meaning of the above terms in this invention according to the specific circumstances. Furthermore, the technical features involved in the different embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0044] See Figure 1 This application provides a surface defect identification system based on multi-angle scattering light field, which can acquire hemispherical spatial scattering information through a single exposure. Its overall structure mainly consists of an illumination system, a sample stage, a hemispherical dome, an aspherical mirror, and a relay optical imaging system.

[0045] The illumination system uses a collimated laser source. The reflective optical element under test is placed on the sample stage at the center of the hemisphere to ensure that the relay optical imaging system can receive sufficient energy. The hemispherical dome is a reflective structure coated with diffuse reflective paint to reflect the multi-angle scattered light from the sample surface in the hemispherical space. The reflected scattered light is reflected again by an aspherical mirror back into the relay optical imaging system, where it is imaged onto the detector. The state of the scratch is inverted by analyzing the light field distribution map.

[0046] The incident luminous flux of the illumination optical system is According to the zenith angle The irradiance received by the sample surface when incident on it is:

[0047] (1)

[0048] In the formula: This indicates the radius of the incident beam.

[0049] After the light source illuminates the sample surface, the total scattered light flux into the hemispherical space is: Radiance can be expressed as:

[0050] (2)

[0051] In the formula: The unit solid angle in a certain direction. The incident angle and azimuth angle of the incident light.

[0052] After passing through the optical system, the detector ultimately receives the radiance distribution from the hemispherical space as follows: ,but:

[0053] (3)

[0054] In the formula, The dome scattering coefficient is... To measure the expression of the scattered light field function of the optical system, , , , These represent the imaging optical system aperture, field of view, focal length, and optical system resolution, respectively. The radiance received by the detector is proportional to the irradiance, and the light intensity distribution of the image directly reflects the multi-angle distribution characteristics of the spatial scattered light field, which will be used for the quantitative analysis and inversion of defects.

[0055] In the surface inspection of optical components, microscopic defects (such as scratches, pits, and bubbles) can disrupt the originally uniform and continuous surface microstructure, thereby altering the reflection behavior of incident light. An ideal optical surface is typically influenced by both specular reflection and Lambertian scattering, and its distribution is stable and directional, which can be described by a bidirectional reflectance distribution function (BRDF). The mathematical definition of BRDF is as follows:

[0056] (4)

[0057] In the formula: The reflection angle and azimuth angle of the reflected light. for Reflected radiance in direction This indicates the surface being measured to receive incident light. The reflectivity, The incident irradiance per unit area, in units of When a defect exists on the surface being measured, the direction of its locally reflected light will change. This manifests as an increase, decrease, or broadening of radiance at certain angles. This change in the spatial distribution of the BRDF (Brake Radiant Fiber Dispersion) can directly reflect the type, size, and other characteristics of the defect, and therefore can serve as an important basis for feature identification and inversion.

[0058] To quantitatively describe this spatial distribution and the detection process, the irradiance received on the detector surface is considered equivalent to the image brightness. The image brightness is proportional to the irradiance, with a proportionality constant of... The absolute measurement expression for the multi-angle scattered light field is obtained as follows:

[0059] (5)

[0060] For a fixed light source with good uniformity, stability, and collimation, the incident illuminance of the light source can be... Treat as a constant The measurement principle expression of BRDF can be converted to:

[0061] (6)

[0062] In the formula, given that the light source, dome scattering coefficient, and measurement optical system are determined, let:

[0063] (7)

[0064] Equation (6) simplifies to Equation (8):

[0065] (8)

[0066] In the formula If is a constant, then the multi-angle reflection distribution function is consistent with the light intensity distribution received by the detector. The surface defects of the tested component are retrieved by analyzing the light intensity distribution received by the detector using an algorithm.

[0067] Different surface defects exhibit varying light intensity distributions in scattered light field images, characterized by features such as the size and symmetry of brightness concentration areas and energy diffusion patterns. These distribution patterns map the differences in the physical structure of the defects and possess good class discrimination capabilities. Therefore, this application provides an optical defect identification method based on multi-angle scattering measurement and deep learning recognition. Light emitted from a laser source strikes the sample under test. Surface defects on the sample cause the reflected light to scatter at multiple angles, which is then captured by a dome structure and a receiving system to form a scattered light field image. This image serves as input to a neural network, extracting spatial light intensity distribution features and outputting defect classification results, thus completing the detection path from physical measurement to identification.

[0068] The deep learning algorithm used in this application is based on the YOLOv11 architecture, constructing an end-to-end defect classification model. The overall structure comprises four core modules: an input layer, a feature extraction module, a C2PSA (Concentration-to-Person Alignment) module, and a classification output module (Classify). The input layer performs image normalization, the feature extraction module captures spatial patterns in the light intensity distribution, the C2PSA module focuses on salient regions to improve discriminative power, and the classification head outputs the defect category. This approach is designed for scattered light field images, which exhibit strong spatial structure regularity and high sensitivity of local features to defect categories.

[0069] To address the challenges of identifying complex defect morphologies, variable spot distributions, and blurred category boundaries in scattered light field images, this application proposes an improved classification and recognition network, SFD-YOLO (Scattering Field Defect YOLO), based on the lightweight backbone structure of YOLOv11 and through structural optimization. Its structure is as follows: Figure 2 As shown.

[0070] SFD-YOLO consists of an input layer, a backbone feature extraction module, an augmented perception module (PSimam), an Efficient_Mamba_CSP semantic modeling module, an attention mechanism (C2PSA), and a classification head. Compared to the original YOLOv11 classification model, SFD-YOLO introduces new modules to enhance network performance in two aspects: shallow feature extraction and deep semantic modeling. First, the PSiam attention module is added in the shallow layer to improve the perception of local salient regions. Second, the Efficient_Mamba_CSP module is used in the deep layer to enhance global feature modeling capabilities. Through these two improvements, SFD-YOLO can more effectively identify defect features in scattered light field images while maintaining the model's lightweight and high efficiency. Specifically, the PSiam module replaces the original CBS module, and the Efficient_Mamba_CSP module replaces the original C3K2 part.

[0071] Shallow Layer Improvements: The PSIram attention module is introduced. The original YOLOv11 backbone's shallow layers contain several convolutional blocks (CBS modules) for extracting low-level features. This paper replaces the two CBS modules in the latter part of the shallow layers with custom PSIram modules while keeping the first two convolutional layers unchanged, to enhance the ability to extract features from locally salient regions. Its structure is as follows: Figure 3As shown, this module consists of multi-scale convolutional paths and the SimAM attention mechanism. First, it extracts spatial features at different scales in parallel, and then integrates them along the channel dimension before introducing them into the SimAM module to achieve parameter-free modeling of spatial saliency. SimAM calculates the importance score of each pixel based on neuron response energy and generates an attention weight map, thereby improving the model's responsiveness to key regions while effectively suppressing background redundancy.

[0072] In-depth improvements: Introducing the Efficient_Mamba_CSP module to replace the C3K2 structure. To enhance YOLOv11's ability to model long-range dependencies in deeper layers, this paper replaces the two C3K2 modules cascaded at the end of the original network with two custom Efficient_Mamba_CSP modules. This module is based on the HSM-SSD (HiddenState Mixer based State Space Duality) mechanism proposed in the EfficientViM architecture, with the core unit being the Efficient_Mamba_block, such as... Figure 4 As shown, it integrates local perception and global state modeling capabilities.

[0073] The Efficient_Mamba_block extracts local spatial features using Depthily Separable Convolutions (DWConv) and introduces an HSM-SSD module for hidden state modeling, mapping feature sequences to a compressed state space to capture long-range dependencies. HSM-SSD offers linear complexity, significantly reducing computational cost compared to self-attention mechanisms. The module also incorporates normalization and residual connections to improve feature stability and gradient propagation efficiency.

[0074] The Efficient_Mamba_CSP module built on this basis, such as Figure 5 As shown, a CSPNet architecture is adopted, dividing the input channels into a main branch and residual branches: the main branch extracts semantic features through the Efficient_Mamba_block, while the residual branches directly transmit the original information. Finally, the input channels are concatenated and fused for output. This design balances deep context modeling with structural efficiency, effectively controlling the consumption of computational resources.

[0075] Compared to the original YOLOv11 classification architecture, this application adopts a partial replacement strategy, introducing only the PSIMam and Efficient_Mamba_CSP modules in the later stages of the network, avoiding significant changes to the YOLOv11 backbone architecture. This design fully preserves the efficiency of the original network in shallow convolutions and mid-level multi-scale modeling, while strengthening local saliency modeling and long-range dependency representation capabilities at key layers. It achieves a balance between accuracy improvement and computational cost control, effectively enhancing the model's feature discrimination and classification robustness in complex scattering images.

[0076] Methods, experiments and results analysis

[0077] 1. Generation of Defect Scattering Light Field Image Dataset

[0078] To verify the proposed method's ability to classify and identify optical surface defects, it is first necessary to establish a typical defect model consistent with actual optical processing scenarios and acquire large-scale multi-angle scattered light field distribution data suitable for deep learning training. This application selects scratch defects, the most common type of defect in optical manufacturing, as the research object. According to international optical component standards (such as ISO 10110-7), scratch width is the core parameter for scratch defect classification and quality grade division, and also the most representative indicator in surface inspection and quality control. Therefore, accurate inversion of the width can not only verify the feasibility of the proposed scattering detection method but also directly serve quality evaluation in actual production.

[0079] First, an adjustable parametric modeling scheme was established for scratch defects. The scratch length was set to 6 mm, the depth to 0.1 mm (matching typical minor processing defects on actual optical surfaces), and the width was sampled at equal intervals (0.1 mm increments) from 0.1 mm to 1.0 mm, totaling 10 specifications, covering the typical grade range from micro-scratches to larger scratches. A defect with a scratch width of 0.2 mm is illustrated below. Figure 6 As shown.

[0080] 2. Experimental parameter settings

[0081] To ensure that the simulation results are consistent with the real-world process, the simulation parameters are combined with common optical materials and typical industrial testing conditions. The specific experimental parameter settings are as follows:

[0082] K9 glass was chosen as the substrate material, as its surface roughness is much smaller than the scratch size and can be ignored; a wavelength of [wavelength value missing] was used. The sample surface is illuminated by a Gaussian light source, with a collimated incident angle. Spot diameter The incident direction is perpendicular to the scratch length; the dome detection surface is divided into 60*60 grids, with a spatial resolution of The number of ray tracing lines was dynamically adjusted for each experiment (from 2 million to 3 million) to ensure that the peak value of the final energy error estimate was below [a certain value]. .

[0083] 3. Dataset Generation

[0084] (1) Generate the scattered light field image dataset required for training and testing according to the following steps: Basic data generation: For each class of width At the center location, 100 sets of light intensity distribution data are generated by randomly adjusting the number of light seeds;

[0085] (2) Enhanced data generation: For to The scratch category, superimposed with positional offset perturbation, from the center along The axis shifts in the opposite direction to Then, the number of light seeds is randomly adjusted to generate 10 sets of light intensity distribution data, and so on. Ten sets of light intensity distribution data were also generated along the positive axis. The sum of the enhanced data center position and the offset data (total number of samples) Group).

[0086] (3) Processing large-width samples: For The scratches were examined, and 100 sets of center position data were kept for comparison with other data. See Table 1 below for details.

[0087]

[0088] The scattered light field images obtained in the experiment were two-dimensional grid data. To improve network training performance, data preprocessing was performed, using standard normalization on the data matrix and bicubic interpolation to uniformly scale the images to a standard size (224×224) to adapt to the input requirements of mainstream convolutional networks. Please refer to [link to relevant documentation]. Figure 7 , Figure 7 The image demonstrates a comparison of typical scattered light field images before and after preprocessing. The left image shows the original 60×60 grid light intensity distribution image, while the right image shows a standard-sized training sample generated after normalization and interpolation. It is evident that the image's detailed structure is effectively preserved during scaling, while the overall brightness distribution is smoother, which helps the subsequent model to stably learn defect features.

[0089] 4. Experimental setup and evaluation indicators

[0090] The hardware configuration used in the experiment is shown in Table 2.

[0091]

[0092] Regarding model training, the parameter configuration is as follows:

[0093] (1) Input image size: 640*640;

[0094] (2) Training rounds: 200 complete iterations are performed on the training dataset;

[0095] (3) Batch-Size: 16;

[0096] (4) Optimizer: Stochastic Gradient Descent (SGD).

[0097] To ensure the accuracy of the experimental results, all experiments were conducted without using pre-trained weights. To objectively evaluate the performance of the SFD-YOLO algorithm, this paper uses the following three mainstream metrics for comparative analysis:

[0098] (1) Top_1 ACC: The most commonly used accuracy metric in classification tasks, representing the proportion of the highest probability class in the prediction results that matches the true label;

[0099] (2) Params: Reflects the complexity and lightweight nature of the model, measured in M;

[0100] (3) Flops: Measure the computational cost of the model, in G units.

[0101] These three metrics allow for a comprehensive assessment of the balance between the model's accuracy, efficiency, and complexity.

[0102] 5. Ablation Experiment Analysis

[0103] To verify that the proposed SFD-YOLO model can be effectively used for the classification and recognition of defect scattering light field images, ablation experiments were conducted. The original YOLOv11 network and the improved network were compared and tested on the test set. The results of the ablation experiments are shown in Table 3.

[0104]

[0105] As shown in Table 3, when only the PSIMam module is introduced, the model's Top-1 ACC is improved to 93.8%, which is 1.3% higher than the original YOLOv11-cls. At the same time, the number of parameters decreases from 1.54M to 1.49M, and the number of Flops increases slightly to 3.6G. This indicates that the shallow attention enhancement module effectively improves the model's ability to capture salient features and improves performance with almost no increase in computational burden.

[0106] When only the Efficient_Mamba_CSP module is introduced, the model's Top-1 ACC reaches 95.1%, accuracy improves by 2.6%, while the number of parameters significantly decreases to 1.30M and Flops remain at 3.6G. This indicates that the module has good lightweight characteristics while improving global modeling capabilities and expression efficiency.

[0107] When both modules are introduced simultaneously to form the complete SFD-YOLO architecture, the Top-1 ACC is further improved to 95.6%, a 3.1% improvement over the original model. The number of parameters is reduced to 1.25M, and the Flops are 4.0G. This demonstrates that while maintaining reasonable computational overhead, this method outperforms YOLOv11-cls in both accuracy and structural efficiency. The results of ablation experiments clearly verify the effectiveness of the improved algorithm.

[0108] like Figure 8 As shown, the accuracy of the SFD-YOLO model gradually increases with each epoch during training, exhibiting a steady upward trend. It converges after approximately 160 epochs, ultimately stabilizing at over 95% accuracy. This result fully demonstrates that the proposed model structure and training strategy possess excellent convergence and optimization performance.

[0109] like Figure 9 As shown, Figure 9 The results show partial predictions for the classification of scattered light field images on the test set. Scattered light field images with widths of 0.1mm, 0.2mm, and 0.4mm all achieved a precision of 1, accurately identifying the images. The discrimination rate for a scattered light field image with a width of 1.0mm was 0.78. This demonstrates that SFD-YOLO can output class predictions consistent with the actual labels for scattering distributions under different defect width parameters. Especially for samples with similar shapes or irregular scattering intensity distributions, while there is a slight deviation at larger widths, it still exhibits stable discrimination ability. These results further validate that the model still possesses good anti-interference performance and discrimination ability when dealing with samples with highly similar categories and low non-local feature significance.

[0110] The visualization of the SFD-YOLO model's classification results for typical defect scattering light field images shows that each sub-image represents the network's classification output for a specific scattering image, with the predicted true class label in the upper left corner. Image colors represent scattering intensity distribution, with red indicating high-intensity areas and blue indicating low-intensity areas. Scattering patterns corresponding to different defect widths (e.g., width_0.3, width_0.4, width_0.8, width_1.0) exhibit significant differences in brightness concentration, shape symmetry, and energy diffusion range. In the 16 sets of scattering light field image recognition results, even though 14 sets of images showed high similarity in light intensity distribution, the model could still accurately classify and identify them, outputting the corresponding predicted label in the upper left corner. The results demonstrate that this model, through its deep global modeling capabilities and shallow attention mechanism, effectively captures differences in key regions of images, achieving reliable discrimination of subtle structural features.

[0111] In summary, the visualization results further validate the stability and application potential of the SFD-YOLO model in multi-angle scattering light field recognition tasks, demonstrating its ability to accurately and automatically classify and scale-invert complex optical defects.

[0112] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0113] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A surface defect identification system based on multi-angle scattering light field, characterized in that, It includes a lighting system, sample stage, hemispherical dome, aspherical reflector, relay optical imaging system and detector, hardware unit and software unit; The illumination system is a collimated laser source, which outputs monochromatic laser light and incident it onto the surface of the optical element under test at a preset zenith angle, with the incident direction perpendicular to the defect length direction. The sample stage is located at the geometric center of the hemispherical dome and is used to fix the optical element to be tested. The inner wall of the hemispherical dome is coated with a diffuse reflection paint to reflect the hemispherical spatial light scattered from the surface of the optical element under test. The aspherical mirror is located on the outside of the hemispherical dome and is used to reflect the scattered light reflected from the dome to the relay optical imaging system. The relay optical imaging system includes a lens group and an aperture, which are used to focus the scattered light onto the detector target surface. The detector is a CCD or CMOS sensor, used to convert the scattered light radiance distribution into a digital scattered light field image; The hardware unit includes a Linux operating system computer, an NVIDIA GPU, an Intel CPU, and supports the PyTorch deep learning framework. The software unit includes a data preprocessing module and an SFD-YOLO network; the data preprocessing module is used to normalize and size-standardize the scattered light field image; the SFD-YOLO network is used to extract features from the scattered image and output defect categories.

2. The system according to claim 1, characterized in that, The SFD-YOLO network is based on YOLOv11-cls architecture and includes an input layer, a backbone feature extraction module, a PSInm attention module, an Efficient_Mamba_CSP module, a C2PSA module, and a classification head. The backbone feature extraction module retains the first two CBS convolutional blocks of YOLOv11-cls, and replaces the two CBS modules in the later shallow layers with PSImam attention modules. The PSImam attention module includes two parallel multi-scale convolutional branches and a SimAM parameterless attention unit. The two deep C3K2 modules of the backbone feature extraction module are replaced with the Efficient_Mamba_CSP module. The Efficient_Mamba_CSP module is based on the CSPNet structure. The main branch includes depthwise separable convolution and HSM-SSD mechanism, and the residual branches directly pass the original features to enhance global semantic modeling.

3. A method for identifying surface defects based on multi-angle scattering light fields as described in any one of claims 1-2, characterized in that, Includes the following steps: S1: Multi-angle scattered light field image acquisition: Adjust the zenith angle of the illumination system, turn on the detector for a single exposure, and acquire the hemispherical spatial scattered light field image of the optical element under test; S2: Dataset generation and preprocessing: Parametrically construct the scratch defect model and generate 1180 sets of samples; normalize the samples and scale them to 640×640 pixels using bicubic interpolation; S3: SFD-YOLO network training: Configure training parameters and perform end-to-end training based on the dataset from step S2 until the model converges; S4: Defect recognition: After preprocessing the image to be detected in step S2, input it into the trained SFD-YOLO network to output the defect category.

4. The method according to claim 3, characterized in that, The dataset generation in step S2 includes: generating 100 samples at the center position and 10 samples each for positive and negative X-axis offsets for scratches with a width of 0.1~0.9mm; generating 100 samples at the center position for scratches with a width of 1.0mm; the number of ray tracing lines for the samples is 2 million to 3 million, and the peak energy error is <5%.

5. The method according to claim 3, characterized in that, In step S3, Efficient_Mamba_block extracts local features through depthwise separable convolution, and maps the feature sequence to the compressed state space through the HSM-SSD mechanism to capture long-distance dependencies with linear computational complexity.

Images