Silica gel component screening and analyzing system based on image recognition

Abstract

The invention relates to a silica gel component screening and analyzing system based on image recognition, and belongs to the cross technical field of image recognition and material analysis. The system comprises a multi-scale microscopic image acquisition unit, a global context perception feature extraction unit, a weak feature enhancement and noise suppression unit, a component semantic segmentation unit, a quantitative analysis and confidence assessment unit and an adaptive threshold decision unit. The problem that imaging of tiny impurities is incomplete under a single magnification is effectively solved, and it is ensured that nanoscale silicon oxide can obtain sufficient characterization under different scales; the global context perception feature extraction module abandons local limitation of traditional convolution, long-range dependence is established in the horizontal direction and the vertical direction through an axial attention mechanism, the model can perceive spatial correlation of weak impurities in the whole view field, and the recognition capacity for isolated and low-contrast particles is remarkably improved.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of image recognition and materials analysis, and specifically relates to a silicone composition screening and analysis system based on image recognition. Background Technology

[0002] In the fields of materials science and industrial quality inspection, image recognition-based component analysis technology has become an important means to improve detection efficiency and accuracy. With the increasing demand for intelligent manufacturing and high-purity materials, higher requirements are being placed on the rapid and accurate identification of trace impurities in composite materials such as silica gel. This technology typically relies on computer vision and deep learning models to achieve automated screening and quantitative analysis of specific components through feature extraction and classification of material microscopic images.

[0003] Image recognition-based silica composition screening analysis focuses on accurately locating and identifying minute impurities such as nanoscale silica from high-resolution microscopic images. Its core objective is to enhance weak feature signals through algorithms, overcoming information loss caused by factors such as small scale and low contrast during imaging, thereby achieving high recall and low false negative rates.

[0004] Significant limitations remain when handling such tasks: mainstream convolutional neural networks are limited by their local receptive fields, making it difficult to effectively capture the contextual relationships of minute impurities in the global image, resulting in insufficient representation capabilities for weak feature components such as nanoscale silica; simultaneously, the visual features of minute components in standard images are easily submerged by background noise, further exacerbating the model's missed detection problem. These issues are particularly prominent in applications with extremely stringent impurity control, such as high-purity silica gel production, severely restricting the reliability and practicality of automated quality inspection systems. Summary of the Invention

[0005] The purpose of this invention is to provide a silicone component screening and analysis system based on image recognition, so as to solve the problems of high false negative rate and low recall rate caused by the limited local receptive field of convolutional neural networks and the easy submersion of weak impurity features by noise in the prior art, thereby realizing high-precision and robust automatic identification and quantitative analysis of trace impurity components such as nano-sized silicon oxide in silicone materials.

[0006] The technical solution of the present invention includes: The multi-scale microscopic image acquisition module is used to acquire high-resolution microscopic image sequences of silicone samples at different magnifications, and to register and fuse the image sequences to generate full-scale fused images with enhanced contrast and detail preservation capabilities. The global context-aware feature extraction module is used to encode features in the full-scale fused image. This module adopts a non-local feature interaction structure based on the axial attention mechanism to build long-distance dependencies along the horizontal and vertical directions to capture the spatial distribution context information of tiny impurities in the whole image. The weak feature enhancement and noise suppression module receives the feature map output by the global context-aware feature extraction module. It dynamically adjusts the response weights of each channel and spatial position through a channel-space dual gating mechanism to suppress irrelevant activation in the background region and amplify the weak signal response in the suspected impurity region. The component semantic segmentation module is used to perform pixel-level semantic segmentation based on the feature map after weak feature enhancement, accurately delineate the spatial contour of nanoscale silicon oxide impurities, and output the corresponding binary mask map. The quantitative analysis and confidence assessment module is used to calculate the number, area ratio, average particle size and spatial distribution density of impurity particles based on the binary mask image generated by the component semantic segmentation module, and to generate a component confidence score for each particle by combining the activation intensity and context consistency index during the segmentation process. The adaptive threshold decision unit is used to dynamically set the impurity judgment threshold based on the confidence score distribution output by the quantitative analysis and confidence assessment module. Particles with a confidence score higher than the threshold are marked as effective silicon oxide components, while those with a confidence score lower than the threshold are regarded as noise or artifacts and are rejected.

[0007] Furthermore, the multi-scale microscopic image acquisition module is equipped with at least three objectives with different magnifications, namely 50x, 100x, and 200x. During the acquisition process, the module synchronously records the illumination intensity, exposure time, and focal plane position parameters of each frame of the image, and uses a sub-pixel-level registration algorithm based on phase correlation to spatially align the multi-scale images. Then, a single full-scale fused image is generated through a Laplacian pyramid fusion strategy, in which high-frequency details are dominated by the high-magnification image and low-frequency structures are dominated by the low-magnification image.

[0008] Furthermore, the global context-aware feature extraction module adopts a four-level downsampling encoder structure, with each encoder level followed by an axial attention block. The axial attention block first performs one-dimensional global attention calculation along the image width direction to generate row-direction context features, then performs one-dimensional global attention calculation along the height direction to generate column-direction context features, and finally performs weighted fusion of the original features, row-direction context features, and column-direction context features. The attention weights are generated through a learnable query-key matching function, which adopts a scaled dot product form, and its scaling factor is the reciprocal of the square root of the feature dimension.

[0009] Furthermore, the weak feature enhancement and noise suppression module includes a channel attention submodule and a spatial attention submodule. The channel attention submodule performs global average pooling along the spatial dimension on the input feature map to obtain the channel description vector, and then generates channel weights through two fully connected layers and a sigmoid activation function. The spatial attention submodule performs max pooling and average pooling along the channel dimension after channel compression, and generates a spatial weight map by concatenating the obtained features and passing them through a 7×7 convolution kernel. The final output feature is the element-wise product of the input feature, the channel weights, and the spatial weights.

[0010] Furthermore, the component semantic segmentation module adopts a U-shaped encoding and decoding architecture. Its encoding path reuses the four-level feature output of the global context-aware feature extraction module, and the decoding path is upsampled through transposed convolution. Skip connections are introduced at each upsampling level to fuse the features of the corresponding encoding level. At the end of the decoder, a dual output head is set. The first output head generates an impurity probability map, and the second output head generates a boundary sharpening map. The two are weighted and fused to generate the final segmentation result. The boundary sharpening map is used to enhance the gradient response of impurity edges and improve contour closure.

[0011] Furthermore, when calculating the confidence level of a single-particle impurity, the quantitative analysis and confidence assessment module integrates three indicators: first, the average activation value of all pixels within the particle region in the probability map output by the component semantic segmentation module; second, the local contrast of the particle in the original full-scale fused image, defined as the absolute value of the difference between the mean gray level inside the particle and the mean gray level of the surrounding annular background region; and third, the consistency score of the particle in multi-scale images, i.e., the percentage of times it is detected in images at 50x, 100x, and 200x magnification. The three indicators are normalized and then weighted and summed, with weight coefficients of 0.5, 0.3, and 0.2, respectively.

[0012] Furthermore, the adaptive threshold decision unit uses a bimodal fitting method based on a Gaussian mixture model to model the distribution of confidence scores for all particles; the component with the lower mean among the two fitted Gaussian components is identified as a noise component, and the component with the higher mean is identified as a real impurity component; the impurity determination threshold is set as the confidence value at the intersection of the two components; if the Gaussian mixture model fails to converge, it regresses to a fixed threshold of 0.65 as the determination criterion.

[0013] Furthermore, this system operates in a closed-loop feedback calibration mode; after each batch of sample analysis is completed, the system automatically pushes particle images with confidence levels between 0.6 and 0.7 to the manual review interface; the review results are used as newly labeled data to fine-tune the decoder parameters of the component semantic segmentation module. The fine-tuning adopts an online incremental learning strategy, updating only the last two decoding layers, with a learning rate of 1 / 10 of that in the initial training phase.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention effectively solves the problem of incomplete imaging of tiny impurities at a single magnification by constructing a multi-scale microscopic image acquisition and fusion mechanism, ensuring that nanoscale silicon oxide can be fully characterized at different scales; the global context-aware feature extraction module abandons the locality limitation of traditional convolution and uses the axial attention mechanism to establish long-range dependencies in the horizontal and vertical directions respectively, enabling the model to perceive the spatial correlation of weak impurities in the entire field of view, significantly improving the ability to identify isolated, low-contrast particles; 2. The weak feature enhancement and noise suppression module achieves active suppression of background noise and directional amplification of impurity signals through a channel-space dual gating mechanism, fundamentally improving the signal-to-noise ratio; the component semantic segmentation module introduces a boundary sharpening output head, which effectively overcomes the edge blurring defect of traditional segmentation models and ensures the geometric accuracy of subsequent quantitative analysis. 3. The quantitative analysis and confidence assessment module integrates multi-dimensional evidence to generate granular confidence levels, avoiding the arbitrariness of a single threshold judgment; the adaptive threshold decision unit dynamically distinguishes noise from real signals based on a Gaussian mixture model, enabling the system to have adaptive robustness to different batches and imaging conditions; the closed-loop feedback calibration mechanism ensures that the system continuously optimizes its performance during long-term operation, forming the ability to self-evolve.

[0015] 4. This invention significantly reduces the false negative rate while ensuring a high recall rate, providing a reliable, traceable, and evolving intelligent quality inspection solution for the production of high-purity silicone. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall technical architecture of the silicone component screening and analysis system based on image recognition proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of the global context-aware feature extraction module in this invention; Figure 3 This is a logical flowchart of the multi-scale microscopic image acquisition and fusion stage in this invention; Figure 4 This is a schematic diagram of the channel-space dual gating mechanism of the weak feature enhancement and noise suppression module in this invention; Figure 5 This is a schematic diagram of the collaborative data flow between the component semantic segmentation module and the quantitative analysis and confidence assessment module in this invention. Detailed Implementation

[0017] Example 1: The overall technical architecture of the image recognition-based silica gel composition screening and analysis system proposed in this invention is shown in the attached figure. Figures 1 to 5As shown, the system consists of six core functional units: a multi-scale microscopic image acquisition module, a global context-aware feature extraction module, a weak feature enhancement and noise suppression module, a component semantic segmentation module, a quantitative analysis and confidence assessment module, and an adaptive threshold decision unit. These units are sequentially connected via data flow, forming a complete closed-loop processing chain from raw image input to impurity component determination output. The entire system runs on a high-performance embedded computing platform. The hardware environment includes a high-precision optical microscopic imaging device, a multi-channel image acquisition card, a graphics processor array, and solid-state storage units. The software adopts a modular design approach, ensuring that each functional unit can be independently deployed, debugged, and upgraded.

[0018] First, the multi-scale microscopic image acquisition module is responsible for acquiring high-resolution microscopic image sequences of the silicone sample under test at different magnifications. This module is equipped with three objectives with magnifications of 50x, 100x, and 200x, corresponding to low-magnification macroscopic structure observation, medium-magnification detail capture, and high-magnification nanoscale impurity imaging, respectively. During image acquisition, the system synchronously records the illumination intensity, exposure time, and focal plane position parameters for each frame. This metadata is used for geometric correction and brightness consistency adjustment in the subsequent image registration and fusion stages. To eliminate image misalignment caused by mechanical vibration, thermal drift, or sample movement, this module employs a sub-pixel-level registration algorithm based on phase correlation to spatially align the multi-scale images. Specifically, for any two images from different magnifications, they are first converted to the frequency domain, the cross-power spectrum is calculated, and then the peak response position is obtained through inverse Fourier transform, thus determining the optimal translation vector. This vector is then refined by bilinear interpolation to achieve sub-pixel-precision spatial alignment. The registered multi-scale image sequence is then used to generate a single full-scale fused image through a Laplacian pyramid fusion strategy.

[0019] In this fusion process, low-frequency structural information is primarily dominated by the 50x image due to its wider field of view and superior depth of field, effectively preserving the overall shape. High-frequency details are provided by the 200x image, which, with its higher spatial resolution, clearly reveals the edge textures and internal structures of nanoscale silica particles. The 100x image serves as an intermediate scale, acting as a transition and smoothing element in the fusion weight allocation. The resulting full-scale fused image combines wide field of view coverage with high detail fidelity, laying a high-quality input foundation for subsequent feature extraction. See the appendix for a detailed explanation of the above process logic. Figure 3 As shown.

[0020] Subsequently, the full-scale fused image is fed into a global context-aware feature extraction module. This module employs a four-level downsampling encoder structure. Each encoder level consists of two 3×3 convolutional layers, a batch normalization layer, and a ReLU activation function stacked on top of each other, with max pooling introduced at the end of each level to achieve 2x downsampling. The key innovation lies in the axial attention block following each encoder level. This block constructs a one-dimensional global attention mechanism along both the image width and height directions, thereby capturing long-distance dependencies in two orthogonal dimensions: horizontal and vertical. (Combined with...) Figure 2 As shown, the workflow of the axial attention block is as follows: First, the feature map of the current layer is unfolded into a sequence of row vectors along the width direction, and a query matrix is ​​generated through a learnable linear projection. AND key matrix The two are scaled dot product operations to calculate the attention weights, which are mathematically expressed as follows:

[0021] in, Let be the dimension of the key vector. As a scaling factor, it is used to alleviate the gradient saturation problem caused by excessively large dot product values ​​in high-dimensional spaces; The value matrix is ​​obtained from the original feature map through another set of learnable projections. After this operation, the output is the row-direction context feature, meaning that the representation of each row of pixels has incorporated information from all rows in the entire image. Similarly, the original feature map is unfolded along the height direction, and the above process is repeated to generate the column-direction context feature. Finally, the original features, row-direction context features, and column-direction context features are weighted and fused. The weight coefficients are dynamically adjusted through learnable parameters to ensure that the model can adaptively balance the contribution ratio of local details and global context according to task requirements. After cascaded processing by a four-level encoder and axial attention blocks, the output deep feature map not only retains rich semantic information but also embeds the distribution correlation of tiny impurities in the entire image space, significantly enhancing the perception ability of isolated, low-contrast particles.

[0022] Next, the aforementioned feature maps are input into the weak feature enhancement and noise suppression module. The core objective of this module is to suppress irrelevant activations in the background region while amplifying the weak signal response in suspected impurity regions. Its structure includes a channel attention submodule and a spatial attention submodule, which are connected serially to form a channel-spatial dual gating mechanism. (See attached...) Figure 4As shown, the channel attention submodule first performs global average pooling along the spatial dimensions (height and width) of the input feature map, obtaining a 1×1×C channel description vector (C being the number of channels). This vector reflects the average activation intensity of each channel across the entire spatial range. Subsequently, this vector is passed sequentially through a first fully connected layer (C / 16 neurons), a ReLU activation function, a second fully connected layer (C neurons), and a Sigmoid activation function to generate a channel weight vector between 0 and 1. This weight vector is element-wise multiplied with the original feature map along the channel dimension, enhancing important channels and suppressing redundant channels. Next, the spatial attention submodule operates on the channel-compressed feature map: first, it performs max pooling and average pooling along the channel dimension, obtaining two H×W×1 feature maps, representing the maximum and average responses at each spatial location, respectively; then, these two are concatenated along the channel dimension to form an H×W×2 composite feature map; finally, a 7×7 large convolution kernel is used to convolve this composite feature map, and a Sigmoid activation function is applied to generate a spatial weight map. This weight map reflects the importance of each spatial location, with high-weight regions corresponding to suspected impurity locations. The final output feature is the element-wise product of the original input feature and the channel weights and spatial weights, i.e.:

[0023] in, Input features For channel weights, Spatial weights, This represents element-wise multiplication. Through this dual gating mechanism, the system effectively improves the signal-to-noise ratio, providing highly discriminative feature representations for subsequent accurate segmentation.

[0024] Subsequently, the enhanced feature maps are fed into the component semantic segmentation module. This module employs a U-shaped encoding / decoding architecture, where the encoding path fully reuses the four-level feature maps output by the global context-aware feature extraction module, avoiding redundant computation and ensuring feature consistency. The decoding path upsamples level by level through transposed convolution operations to restore the original image resolution. At each upsampling level, the system introduces skip connections, concatenating the feature maps of the corresponding encoding level with the currently decoded features to fuse shallow detail information with deep semantic information, mitigating the boundary blurring problem caused by multiple downsampling. Crucially, this module features a dual-output head structure at the decoder end: The first output head generates an impurity probability map, where each pixel value represents the probability that the location belongs to a silicon oxide impurity; The second output head generates a boundary sharpening map, which is trained using a specially designed edge-supervised loss function designed to enhance the gradient response at impurity edges.

[0025] The final segmentation result is obtained by weighted fusion of the two methods. The fusion weights are automatically learned during the training phase. Typically, the boundary sharpening map has a higher weight in edge regions, while the probability map dominates in interior regions. This design significantly improves the closure and geometric accuracy of the segmentation contour, ensuring the reliability of subsequent quantitative analysis calculations of parameters such as area and particle size. The collaborative processing flow is attached. Figure 5 As shown.

[0026] After obtaining the accurate binary mask image, the system proceeds to the quantitative analysis and confidence assessment module. This module first performs connected component analysis on the mask image to identify all independent impurity particles and assigns a unique identifier to each particle. For each particle, the system calculates four basic geometric parameters: particle count (i.e., the total number of connected components), total area ratio (the proportion of all particle pixels to the total number of pixels in the entire image), average particle size (calculated based on the equivalent circle diameter), and spatial distribution density (the number of particles per unit area). More importantly, this module generates a component confidence score for each particle, which integrates three independent pieces of evidence: first, the average activation value of all pixels within the particle's region in the probability map output by the component semantic segmentation module, reflecting the model's confidence in the region's discrimination; second, the particle's local contrast in the original full-scale fused image, defined as the absolute value of the difference between the particle's internal grayscale mean and the surrounding annular background region (extending 5 pixels outward), used to measure physical imaging discriminability; and third, the particle's consistency score across multiple scales, i.e., the percentage of times it is detected in all three sets of original images at 50x, 100x, and 200x magnification, used to verify its cross-scale stability. These three indicators are denoted as follows: After min-max normalization to the [0,1] interval, the results are weighted and summed according to preset weights of 0.5, 0.3, and 0.2, respectively. The final confidence score is:

[0027] This scoring mechanism avoids the one-sidedness of a single indicator and enables decision-making based on the fusion of multi-source evidence.

[0028] Finally, the confidence scores of all particles are fed into an adaptive threshold decision unit. This unit uses a bimodal fitting method based on a Gaussian mixture model to model the score distribution. Specifically, it is assumed that the score data is generated by mixing two Gaussian components, corresponding to the noise / artifact population and the real impurity population, respectively. The model parameters are iteratively optimized using the expectation-maximization algorithm until convergence. If the fit is successful, the component with the lower mean is identified as the noise component, and the component with the higher mean is identified as the real impurity component. The impurity determination threshold is set as the confidence value at the intersection of the probability density curves of the two components, which satisfies the condition that the posterior probabilities of the two components are equal, possessing the optimal Bayesian classification boundary property. If the Gaussian mixture model fails to converge due to a unimodal data distribution or insufficient sample size (such as log-likelihood function oscillation or iteration overlimit), the system automatically reverts to a preset fixed threshold of 0.65 as the judgment benchmark, ensuring that the system still has basic discrimination ability under abnormal operating conditions. All particles with confidence scores higher than the threshold are marked as effective silica components and included in the final quality inspection report; those with scores lower than the threshold are considered noise or artifacts and are removed.

[0029] Furthermore, this system operates in a closed-loop feedback calibration mode. After each batch of samples is analyzed, the system automatically filters particle images with a confidence level between 0.6 and 0.7—this range is considered a "fuzzy zone," meaning it is neither a high-confidence true impurity nor significant noise. These images are then pushed to a manual review interface for visual confirmation by quality control experts. The review results (confirmed as impurities or not) are used as new labeled data to fine-tune the decoder parameters of the component semantic segmentation module. Fine-tuning employs an online incremental learning strategy, updating only the last two convolutional layers of the decoder to avoid catastrophic forgetting; the learning rate is set to 1 / 10 of the initial training phase (e.g., decreasing from 0.001 to 0.0001) to ensure smooth and stable parameter updates. This mechanism enables the system to continuously learn and self-evolve, adapting to differences in material composition, imaging conditions, and the emergence of new impurity types in different batches of silicone samples.

[0030] In summary, this embodiment constructs a high-precision and robust intelligent analysis system for silicone composition by synergistically integrating six core technologies: multi-scale imaging fusion, axial attention-driven global context modeling, channel-space dual gating enhancement, boundary-aware segmentation, multi-dimensional confidence assessment, and adaptive threshold decision-making. This system not only solves the problem of missed detections caused by the limited local receptive field of traditional convolutional neural networks, but also achieves long-term performance optimization through a closed-loop feedback mechanism, providing a traceable, interpretable, and evolving quality control tool for high-purity silicone production.

[0031] Example 2: Building upon Example 1, this example further optimizes the illumination control strategy of the multi-scale microscopic image acquisition module to enhance the imaging contrast of nanoscale silica impurities. Specifically, instead of using a fixed illumination intensity, this module introduces a dynamic illumination control mechanism. Before acquiring 50x, 100x, and 200x images, the system first performs a low-resolution pre-scan of the sample to obtain an overall reflectance distribution map. Based on this distribution map, the system divides the image into several brightness regions and independently calculates the optimal illumination intensity for each region. For high reflectance regions (such as silicone substrates), a lower illumination intensity is used to avoid overexposure; for low reflectance regions (such as suspected impurity aggregation areas), the illumination intensity is increased to enhance signal response. This control strategy is implemented using a digital micromirror device with a response time of less than 10 milliseconds, ensuring real-time switching during high-speed imaging.

[0032] Furthermore, this embodiment improves the spatial attention submodule of the weak feature enhancement and noise suppression module. The original 7×7 convolutional kernel is replaced with a deformable convolutional kernel, whose sampling position is dynamically generated by a lightweight offset prediction network. This network takes the input feature map as input and outputs the horizontal and vertical offsets at each spatial location, enabling the convolutional receptive field to adaptively focus on the irregular shapes of impurity edges, rather than fixed rectangular regions. This improvement significantly enhances the accuracy of spatial weight allocation for non-spherical, irregular silicon oxide particles.

[0033] In the component semantic segmentation module, this embodiment introduces a multi-task joint training mechanism. In addition to the main segmentation task, an auxiliary task for impurity type classification is added. Its input is a cropped image of each connected component, and its output is a predefined impurity category (such as silicon oxide, metal oxides, organic residues, etc.). Although this system currently only filters for silicon oxide, this auxiliary task, by sharing encoder features, forces the model to learn a more discriminative general impurity representation, indirectly improving the accuracy of silicon oxide segmentation. During training, the total loss function is a weighted sum of the segmentation loss and the classification loss, with a weight ratio of 4:1.

[0034] The quantitative analysis and confidence assessment module has also been enhanced. In addition to the original three indicators, a fourth indicator has been added: particle morphology regularity. This indicator is calculated by determining the roundness of the particle outline (4π × area / perimeter). 2 The weighted combination of the regularity score and the ellipse fitting residual is used to exclude false positive segments caused by non-component factors such as bubbles and scratches. The regularity score is normalized and added to the final confidence score calculation with a weight of 0.1, so that the total weight distribution is adjusted to 0.5, 0.3, 0.2, and 0.1.

[0035] In this embodiment, the adaptive threshold decision unit supports multimodal Gaussian mixture models. When multiple impurity types exist in the sample, the system can automatically extend to three- or four-peak fitting, selecting the optimal number of components using Bayesian information criteria, thereby achieving differentiated threshold settings for different impurity groups. Although this function is not currently enabled in the single-component screening scenario, it reserves a technical interface for future system expansion.

[0036] The closed-loop feedback calibration mode has also been upgraded to a semi-supervised active learning mechanism. The system not only pushes particles in the 0.6–0.7 range, but also actively selects the most informative samples for manual annotation based on prediction uncertainties (such as Monte Carlo Dropout variance), thereby achieving maximum model gain with minimal human intervention. This mechanism significantly reduces manpower costs in long-term operation and maintenance.

[0037] Through the above improvements, this embodiment, while maintaining the core architecture of Embodiment 1, further enhances the system's adaptability and analysis accuracy in complex lighting, irregular impurity morphology, and multi-impurity coexistence scenarios, verifying the scalability and engineering applicability of the technical route of this invention.

Claims

1. A silicone component screening and analysis system based on image recognition, characterized in that, include: The multi-scale microscopic image acquisition module is used to acquire high-resolution microscopic image sequences of silicone samples at different magnifications, and to register and fuse the image sequences to generate full-scale fused images with enhanced contrast and detail preservation capabilities. The global context-aware feature extraction module is used to encode features in the full-scale fused image. This module adopts a non-local feature interaction structure based on the axial attention mechanism to build long-distance dependencies along the horizontal and vertical directions to capture the spatial distribution context information of tiny impurities in the whole image. The weak feature enhancement and noise suppression module receives the feature map output by the global context-aware feature extraction module. It dynamically adjusts the response weights of each channel and spatial position through a channel-space dual gating mechanism to suppress irrelevant activation in the background region and amplify the weak signal response in the suspected impurity region. The component semantic segmentation module is used to perform pixel-level semantic segmentation based on the feature map after weak feature enhancement, accurately delineate the spatial contour of nanoscale silicon oxide impurities, and output the corresponding binary mask map. The quantitative analysis and confidence assessment module is used to calculate the number, area ratio, average particle size and spatial distribution density of impurity particles based on the binary mask image generated by the component semantic segmentation module, and to generate a component confidence score for each particle by combining the activation intensity and context consistency index during the segmentation process. The adaptive threshold decision unit is used to dynamically set the impurity judgment threshold based on the confidence score distribution output by the quantitative analysis and confidence assessment module. Particles with a confidence score higher than the threshold are marked as effective silicon oxide components, while those with a confidence score lower than the threshold are regarded as noise or artifacts and are rejected.

2. The silicone composition screening and analysis system based on image recognition according to claim 1, characterized in that, The multi-scale microscopic image acquisition module is equipped with at least three objectives with different magnifications, namely 50x, 100x, and 200x. During the acquisition process, the module synchronously records the illumination intensity, exposure time, and focal plane position parameters of each frame of the image. It then uses a sub-pixel-level registration algorithm based on phase correlation to spatially align the multi-scale images and generates a single full-scale fused image through a Laplacian pyramid fusion strategy. High-frequency details are dominated by the 200x image, while low-frequency structures are dominated by the 50x image.

3. The silicone component screening and analysis system based on image recognition according to claim 1, characterized in that, The global context-aware feature extraction module adopts a four-level downsampling encoder structure, with an axial attention block following each encoder level. The axial attention block first performs one-dimensional global attention calculation along the image width direction to generate row-direction context features, then performs one-dimensional global attention calculation along the height direction to generate column-direction context features, and finally performs weighted fusion of the original features, row-direction context features and column-direction context features. The attention weights are generated through a learnable query-key matching function, which is a scaled dot product with a scaling factor that is the reciprocal of the square root of the feature dimension.

4. The silicone component screening and analysis system based on image recognition according to claim 1, characterized in that, The weak feature enhancement and noise suppression module includes a channel attention submodule and a spatial attention submodule. The channel attention submodule performs global average pooling on the input feature map along the spatial dimension to obtain the channel description vector, and then generates channel weights through two fully connected layers and a sigmoid activation function. The spatial attention submodule performs max pooling and average pooling along the channel dimension after channel compression, and then generates a spatial weight map by concatenating the resulting features and passing them through a 7×7 convolution kernel. The final output feature is the element-wise product of the input feature and the channel weights and spatial weights.

5. The silicone component screening and analysis system based on image recognition according to claim 1, characterized in that, The component semantic segmentation module adopts a U-shaped encoding and decoding architecture. Its encoding path reuses the four-level feature output of the global context-aware feature extraction module, and the decoding path is upsampled through transposed convolution. Skip connections are introduced at each upsampling level to fuse the features of the corresponding encoding level. A dual output head is set at the end of the decoder. The first output head generates an impurity probability map, and the second output head generates a boundary sharpening map. The two are weighted and fused to generate the final segmentation result. The boundary sharpening map is used to enhance the gradient response of impurity edges and improve contour closure.

6. The image recognition-based silica gel composition screening and analysis system according to claim 1, characterized in that, The quantitative analysis and confidence assessment module considers three indicators when calculating the confidence level of a single-particle impurity: One is the average activation value of all pixels within the particle region in the output probability map of the component semantic segmentation module; The second is the local contrast of the particle in the original full-scale fused image, which is defined as the absolute value of the difference between the mean gray value inside the particle and the mean gray value of the surrounding annular background area. The third is the consistency score of the particle in multi-scale images, that is, the percentage of times it is detected in images at 50x, 100x and 200x magnification. The three indicators were normalized and then weighted and summed, with weight coefficients of 0.5, 0.3 and 0.2 respectively.

7. The silicone composition screening and analysis system based on image recognition according to claim 1, characterized in that, The adaptive threshold decision unit uses a bimodal fitting method based on a Gaussian mixture model to model the distribution of confidence scores for all particles. The component with the lower mean among the two Gaussian components obtained from the fitting is identified as the noise component, and the component with the higher mean is identified as the real impurity component. The impurity judgment threshold is set as the confidence value at the intersection of the two components; if the Gaussian mixture model fails to converge, it will fall back to the fixed threshold of 0.65 as the judgment criterion.

8. The silicone component screening and analysis system based on image recognition according to claim 1, characterized in that, The system operates in a closed-loop feedback calibration mode; after each batch of sample analysis is completed, the system automatically pushes particle images with a confidence level between 0.6 and 0.7 to the manual review interface. The verification results are used as newly labeled data to fine-tune the decoder parameters of the component semantic segmentation module. The fine-tuning adopts an online incremental learning strategy, updating only the last two decoding layers.

9. The image recognition-based silica gel composition screening and analysis system according to claim 2, characterized in that, The multi-scale microscopic image acquisition module performs a low-resolution pre-scan of the sample before image acquisition to obtain an overall reflectance distribution map. Based on this distribution map, the image is divided into several brightness regions, and the optimal illumination intensity is calculated independently for each region. Real-time illumination control is achieved through a digital micromirror device.

10. The silicone component screening and analysis system based on image recognition according to claim 4, characterized in that, In the spatial attention submodule of the weak feature enhancement and noise suppression module, the 7×7 convolutional kernel is replaced with a deformable convolutional kernel. Its sampling position is dynamically generated by a lightweight offset prediction network. This network takes the input feature map as input and outputs the horizontal and vertical offsets of each spatial position, so that the convolutional receptive field adaptively focuses on the irregular shape of the impurity edge.

Images