Prism film surface defect recognition and classification management system and method

Abstract

The prism film surface defect recognition and classification management system and method, the system comprises a multi-modal imaging module for collecting original image signals; a processing module receives the original image signals, constructs a background model based on prior knowledge of the periodic microstructure of the prism film, and generates pre-processed image signals after suppressing the background; a fusion module receives the pre-processed image signals, synchronously extracts multi-scale defect morphological features and texture features using a small sample optimized deep learning model, and forms a fusion feature signal; a decision module receives the fusion feature signal and outputs a defect type and a potential process correlation factor classification signal; a feedback module receives the factor classification signal and transmits the instruction to the production line control system to trigger dynamic process parameter adjustment. The prism film surface defect recognition and classification management system and method can solve the problem of high surface micro-defect omission rate and inaccurate classification caused by strong light reflection and complex texture interference of the microstructure of the prism film.

Description

Technical Field

[0001] The present invention relates to the technical field of surface quality inspection of optical functional films, and specifically to a prism film surface defect recognition and classification management system and method. Background Art

[0002] As a key brightness enhancement component for liquid crystal displays, processing defects (such as scratches, bubbles, foreign objects) on the surface of the prism film will cause problems such as uneven display and brightness reduction. Current mainstream detections rely on manual visual inspection or traditional machine vision. However, the manual method has defects such as low efficiency and large subjective deviation; although traditional machine vision uses line array scanning or bright and dark field imaging, due to the strong anisotropic reflection generated by the periodic microstructures of the prism film, the high-light areas cover the real defects during imaging, and the texture background interferes with the recognition of small targets. Especially for low-contrast pits and transparent foreign objects, the detection rate of existing algorithms is insufficient. At the same time, the detection technology based on deep learning is limited by the problem of scarce samples - in actual production lines, the vast majority are qualified products, the defect samples are scarce and the forms are variable, making it difficult to train a highly generalized model.

[0003] In terms of defect management, existing systems are mostly limited to classification and counting, lacking in-depth analysis of the causes of defects. The classification results are separated from data such as process parameters and equipment status, and a defect prevention mechanism cannot be established. For example, scratches may be caused by mold damage or insufficient cleaning, but traditional methods cannot distinguish the cause types; bubble defects may be due to temperature fluctuations or water content in raw materials, but they have not been effectively traced. More critically, the detection results are feedback-delayed and cannot adjust production parameters in real time, resulting in continuous generation of defects. The industry urgently needs an intelligent detection system that can overcome strong specular reflection interference, integrate small-sample learning ability, associate physical causes, and achieve closed-loop control. Summary of the Invention

[0004] In view of the above-mentioned disadvantages of the prior art, the purpose of the present invention is to provide a prism film surface defect recognition and classification management system and method, which are used to solve the problems of high missed detection rate and inaccurate classification of small surface defects of the prism film due to strong specular reflection and complex texture interference of the microstructures. The present invention suppresses specular reflection interference through multi-angle light source collaborative imaging, uses the prior knowledge of the prism period for frequency domain filtering to separate the background; adopts a small-sample deep learning model to fuse multi-scale defect features; associates the physical form of the defect with the cause knowledge base to achieve interpretable classification; finally generates a feedback instruction containing process optimization suggestions to dynamically regulate the production line, forming a "detection - analysis - control" closed loop. The present invention provides a prism film surface defect recognition and classification management system, including: A multi-modal imaging module, which collects the original image signals of the prism film surface under different optical conditions by collaboratively controlling multi-angle light sources; The processing module receives the original image signal, constructs a background model based on prior knowledge of the periodic microstructure of the prism film, separates background interference through frequency domain filtering and reflection compensation algorithms, and generates a preprocessed image signal after background suppression. The fusion module receives the preprocessed image signal, uses a small-sample optimized deep learning model to simultaneously extract multi-scale defect morphology features and texture features, and adaptively weights and fuses the feature maps of different scales to form a fused feature signal. The decision module receives the fused feature signal, associates the physical morphological features of the defect with the preset defect cause knowledge base through the classification network, and outputs the defect type and potential process-related factor classification signal. The feedback module receives factor classification signals, generates feedback instructions in real time that include defect type, location coordinates and process optimization suggestions, and transmits the instructions to the production line control system to trigger dynamic process parameter adjustments.

[0005] In one embodiment of the present invention, when the multimodal imaging module coordinates the control of multi-angle light sources, it is specifically configured to dynamically adjust the combination of incident angles and polarization directions of each light source according to the preset spatial frequency distribution of the microstructure on the prism film surface. The combination of incident angles includes coaxial illumination perpendicular to the film surface, low-angle grazing illumination less than ten degrees, and oblique diffuse illumination between thirty and sixty degrees. By triggering image sensors under different illumination modes in a time-division manner, multiple sets of original image signals with complementary defect response characteristics are acquired to ensure the imaging integrity of high reflectivity areas and low contrast defect areas.

[0006] In one embodiment of the present invention, when constructing the background model, the processing module specifically performs Fourier transform analysis on the spatial frequency of the periodic microstructure of the prism film to extract the fundamental frequency and harmonic components as the background template; the frequency domain filtering algorithm selectively attenuates the frequency domain energy corresponding to the background template by designing a band-stop filter, while retaining the abnormal frequency domain components caused by the defects; the reflection compensation algorithm uses the principle of polarization differential imaging to extract and suppress the specular reflection components from mutually orthogonal polarization channel images, and the final preprocessed image signal with improved signal-to-noise ratio contains detextured defect enhancement information.

[0007] In one embodiment of the present invention, the fusion module employs a few-sample optimized deep learning model, which transfers the basic feature extraction layer of the pre-trained convolutional neural network and freezes its underlying weights to retain general texture recognition capabilities; a synthetic defect sample generation mechanism is introduced in the model fine-tuning stage, which simulates the optical scattering model of scratches, bubbles, and foreign object attachment defects based on the physical properties of the prism film, and fuses the generated virtual defect images with the real defect-free background to expand the training samples; a multi-branch feature pyramid structure is deployed at the model output end to extract morphological feature maps and local binary pattern texture feature maps under different receptive fields.

[0008] In one embodiment of the present invention, when the fusion module performs adaptive weighted fusion, it dynamically allocates weight coefficients based on the defect saliency heatmap of each scale feature map; wherein the saliency heatmap is generated by calculating the variance of activation values ​​between feature map channels, and higher weights are given to channels with high variance to enhance the response to minor defects; after the weighted multi-scale feature maps are spliced ​​by channels and dimensionality reduced by one-to-one convolution, a fused feature signal with both local details and global semantics is formed.

[0009] In one embodiment of the present invention, the classification network of the decision module adopts a graph neural network architecture, which maps the fused feature signals to node attributes; the preset defect cause knowledge base is constructed as a heterogeneous knowledge graph, whose nodes include defect types, process parameters, and material attribute entities, and the edge relationships define the physical association rules between defects and causes; the classification process calculates the similarity between the fused feature nodes and the knowledge graph nodes through a graph attention mechanism, and matches the defect type node with the highest similarity and its associated process factor node, thereby realizing the interpretable mapping from physical morphological features to causal factors.

[0010] In one embodiment of the present invention, when the feedback module generates process optimization suggestions, it performs multi-level searches in a preset process adjustment rule library based on the defect type and process correlation factor classification signal. The primary search matches the process parameter sensitivity ranking table corresponding to the defect type, the secondary search filters the top three key process parameters based on the correlation factor strength, and finally combines the clustering results of the defect location coordinates in the production line spatial distribution model to generate parameter tuning instructions for specific equipment.

[0011] In one embodiment of the present invention, a data buffer interface is provided between the processing module and the fusion module. This interface is configured to divide and reassemble the preprocessed image signal into blocks; divide the entire image into overlapping local region blocks; allocate different feature extraction computing resources according to the background complexity score of each region block; enable full-resolution processing for high-complexity regions and downsampling processing for low-complexity regions; and achieve balanced computing load and real-time performance.

[0012] In one embodiment of the present invention, the decision module initiates an online verification mechanism after outputting the classification signal: when there is a conflict in the classification results of the same location under different imaging modalities, the reprocessing process of the multimodal original image signal is triggered; by comparing the confidence distribution of the defect features under each modality, the classification result corresponding to the modality with the highest confidence is selected as the final output, and the conflict cases are automatically added to the incremental training set of the small sample optimization model.

[0013] This invention also includes a method for identifying and classifying defects on the surface of a prism film, comprising: S1: By coordinating the control of multi-angle light sources, the original image signals of the prism film surface under different optical conditions are acquired; S2: Receive the original image signal, construct a background model based on prior knowledge of the periodic microstructure of the prism film, separate background interference through frequency domain filtering and reflection compensation algorithm, and generate a preprocessed image signal after background suppression; S3: Receive the preprocessed image signal, use a small-sample optimized deep learning model to simultaneously extract multi-scale defect morphology features and texture features, and adaptively weight and fuse feature maps of different scales to form a fused feature signal; S4: Receive the fused feature signal, associate the physical morphological features of the defect with the preset defect cause knowledge base through the classification network, and output the defect type and potential process correlation factor classification signal. S5: Receives factor classification signals, generates feedback instructions in real time including defect type, location coordinates and process optimization suggestions, and transmits the instructions to the production line control system to trigger dynamic process parameter adjustments.

[0014] The prism film surface defect identification and classification management system and method provided by this invention suppresses reflective interference through multi-angle light source collaborative imaging, uses prism periodic prior knowledge for frequency domain filtering to separate the background, employs a small-sample deep learning model to fuse multi-scale defect features, associates defect physical morphology with a knowledge base of causes to achieve interpretable classification, and finally generates feedback instructions containing process optimization suggestions to dynamically control the production line, forming a "detection-analysis-control" closed loop. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 System architecture diagram for prism film surface defect identification and classification management system; Figure 2This is a flowchart of a method for identifying and classifying defects on the surface of a prism film. Detailed Implementation

[0017] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0018] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0019] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0020] Please see Figure 1-2 The diagram illustrates the prism film surface defect identification and classification management system and method of the present invention. The prism film surface defect identification and classification management system includes a multimodal imaging module, which acquires original image signals of the prism film surface under different optical conditions by collaboratively controlling multi-angle light sources; a processing module, which receives the original image signals, constructs a background model based on prior knowledge of the periodic microstructure of the prism film, separates background interference through frequency domain filtering and reflection compensation algorithms, and generates a preprocessed image signal after background suppression; a fusion module, which receives the preprocessed image signals, simultaneously extracts multi-scale defect morphological features and texture features using a small-sample optimized deep learning model, and adaptively weights and fuses the feature maps of different scales to form a fused feature signal; a decision module, which receives the fused feature signal, associates the physical morphological features of the defects with a preset defect cause knowledge base through a classification network, and outputs a defect type and potential process-related factor classification signal; and a feedback module, which receives the factor classification signal, generates a feedback instruction in real time containing defect type, location coordinates, and process optimization suggestions, and transmits the instruction to the production line control system to trigger dynamic process parameter adjustments.

[0021] like Figure 1As shown, the system consists of five core modules forming a data closed loop. The multimodal imaging module serves as the physical signal input terminal, integrating multiple programmable light-emitting units and a high-resolution image sensor. The light-emitting units include three basic illumination components: coaxial cold light source, low-angle linear array light source, and oblique diffuse light source. Each component can independently adjust its wavelength, intensity, and polarization state. The imaging control unit automatically generates the optimal illumination combination scheme based on a pre-set microstructure parameter library for the prism film model. For example, coaxial light combined with 30-degree oblique diffuse light is used for high-density prism arrays to suppress specular reflection, while low-angle grazing light is used for wide-spacing prisms to enhance scratch contrast. The image sensor acquires raw image sequences corresponding to different illumination conditions under a time-division triggering mechanism, with each frame image accompanied by optical parameter metadata. After receiving the original image sequence, the processing module first performs frequency domain background modeling: extracting the spatial distribution features of the periodic microstructure of the prism film, converting it to the frequency domain through fast Fourier transform, and identifying the fundamental frequency and harmonic component sets representing the background texture; based on this component set, a two-dimensional band-stop filter is constructed, and convolution operation is performed on the spectrum of the original image to attenuate the background energy, while retaining the abnormal high-frequency and low-frequency residues caused by defects; then the reflection compensation process is started, and the specular reflection difference field between orthogonal polarization channel images is calculated using the polarization difference algorithm, and the difference field is subtracted from the original image to eliminate specular artifacts, finally outputting a preprocessed image with improved signal-to-noise ratio. The fusion module loads a pre-trained few-sample optimized model to process the pre-processed image. This model uses a feature pyramid network as its backbone. In the encoder part, the pre-trained weights are frozen to retain the general texture recognition capability, while in the decoder part, dynamic convolutional layers are introduced to adapt to the prism membrane characteristics. The model extracts feature maps at three scales in parallel: the bottom feature map focuses on local details to capture micron-level point defects, the middle feature map identifies the continuous shape of linear defects, and the high-level feature map resolves the global semantic relationship between defects and the background. Multi-scale features are fused through a saliency-driven adaptive weighting mechanism—the channel attention score of each feature map is calculated, and the high-score channels are given exponentially increasing weight coefficients to strengthen the key defect signals. The weighted feature maps are then concatenated by channels and compressed by one-to-one convolution to generate a fused feature vector. After receiving the fused feature vector, the decision module inputs it into a graph neural network classifier. This classifier is pre-embedded with a defect cause knowledge graph constructed by process experts. The graph nodes contain sixteen standard defect types and their associated process factor entities. The classifier calculates the similarity distribution between the feature vector and each defect type node, matches the defect type label with the highest similarity, and simultaneously traverses the edge relationships of the knowledge graph to retrieve the corresponding set of process-related factors, outputting a structured classification signal.The feedback module analyzes the defect type and process factors in the classification signal, calls the process rule library to generate optimization suggestions, such as polishing repair for mold damage scratches and drying pretreatment for water-containing air bubbles in the raw material; at the same time, it integrates the coordinate information of defects on the film surface to generate a spatial distribution heat map, and pushes the feedback command containing suggestions and heat map to the production line control system in real time through the industrial communication protocol to trigger dynamic compensation of parameters such as extrusion temperature and pressure roller pressure.

[0022] Furthermore, the module's lighting coordination control is based on the optical response characteristics of the prism film microstructure. The incident angle combination generation unit incorporates a microstructure parameter-optical response mapping table, established through offline optical simulation, recording the optimal incident angle range for each lighting mode under different prism heights, apex angles, and arrangement periods. When the system inputs the film model, the control unit retrieves the mapping table to determine the initial lighting scheme. For example, for an equilateral prism array with a 90-degree apex angle, the default is a combination of coaxial light and 45-degree oblique light. Polarization direction control is achieved by a rotating motor driving a linear polarizer, configuring differentiated polarization strategies for different defect types: when detecting surface foreign objects, zero-degree and 90-degree orthogonal polarization are used to suppress background scattering; when detecting internal bubbles, 45-degree and 135-degree polarization are switched to enhance the body scattering signal. The time-division triggering mechanism is managed by a high-precision timing controller. Within a single detection cycle, three sets of illumination modes are activated sequentially. The exposure time of each mode is adaptively adjusted according to the light source intensity to ensure that each modal image has a consistent brightness benchmark. The image sensor synchronously outputs a raw image sequence with timestamps. Pixels at the same position in the sequence are stored as multi-dimensional data cubes according to their illumination mode. To ensure imaging integrity, the system sets up a feedback adjustment loop: after acquiring the first frame image, the proportion of highlight areas and local contrast indicators are calculated in real time. If the highlight area exceeds a set threshold, the coaxial light intensity is reduced and oblique light compensation is increased. If the local contrast is too low, a low-angle light source secondary imaging is activated. The final output raw image signal includes three geometrically aligned multimodal images and corresponding optical parameter logs.

[0023] Specifically, the background modeling stage employs a hybrid strategy combining frequency and spatial domain approaches. First, the input original image is divided into overlapping blocks, with each block undergoing an independent two-dimensional Fourier transform to obtain its amplitude spectrum. A clustering algorithm identifies significant peak points in the amplitude spectra of each block, and peak points with similar spatial frequencies are selected to form a candidate set of background fundamental frequencies. Prior knowledge of the prism period is used to verify the rationality of the candidate set, eliminating outliers deviating more than 20% from the theoretical frequency, ultimately generating the background fundamental frequency template and its harmonic order list. The frequency domain filtering algorithm designs an adaptive band-stop filter bank, with each filter corresponding to a background frequency component: the stopband center frequency is derived from the fundamental frequency template, the stopband width dynamically expands according to harmonic energy, and the stopband attenuation depth is proportional to the background harmonic amplitude. After applying the filter bank to the entire image spectrum, an inverse Fourier transform is performed to reconstruct the spatial image. At this point, background texture is significantly suppressed, but periodic artifacts remain. The reflection compensation algorithm employs a dual-polarization difference process to address artifact issues: Zero-degree and 90-degree polarized images under the same illumination conditions are extracted from multimodal images, and pixel-level differences between the two images are calculated to generate a difference map. High-confidence specular reflection regions are extracted from the difference map through threshold segmentation, and a full-image specular reflection component model is constructed using radial basis function interpolation. This model is subtracted from the original image, and dark current noise is compensated to obtain a preprocessed image with a uniform background. To handle special scenarios, the algorithm incorporates an anomaly recovery mechanism: when frequency domain filtering causes excessive attenuation of the effective defect signal, abnormally steep transition points in the image gradient histogram are detected, and the filter is disabled in the corresponding region, switching to spatial domain morphological background subtraction. The final output preprocessed image meets the technical requirements of improving the signal-to-noise ratio of the defect region by more than three times and reducing the background standard deviation to one-fifth of the original image.

[0024] In one embodiment of the present invention, the fusion module employs a few-shot optimization model architecture, with an improved feature pyramid network as the main framework. During model initialization, a residual network pre-trained on a public defect dataset is loaded as the encoder, and its first three convolutional weights are frozen to retain general edge and texture extraction capabilities. Subsequent convolutional layers are unfrozen, and an adaptive instance normalization module is inserted to inject the statistical characteristics of the prism membrane background into the feature space through affine transformation. The decoder section is designed with three parallel upsampling branches, each corresponding to feature extraction in a different receptive field: the first branch retains a high-resolution feature map (one-quarter the size of the original image) and captures the sub-pixel contours of point defects through dilated convolutions with a dilation rate of one; the second branch downsamples to one-eighth the size and applies long short-term memory convolutional layers to model the directional continuity of scratch-like linear defects; the third branch compresses to one-sixteenth the size and uses global attention pooling to extract the semantic context of the defect region. To address the scarcity of real-world defect samples, a physics-driven synthetic data generator is introduced during the model training phase. This involves constructing an optical refraction model of the foreign particles based on Mie scattering theory, simulating the diffraction pattern of the scratches according to the Fraunhofer diffraction formula, and rendering the scattering light path inside the bubble using the Monte Carlo method. The generated physically simulated defects are then embedded into a defect-free prism film background image, and a Poisson mixture algorithm is used to achieve consistent illumination fusion. The training process employs a two-stage strategy: the first stage uses hundreds of thousands of synthetic images to train the decoder parameters, with the loss function combining the cross-union ratio and perceptual loss; the second stage freezes the decoder, fine-tunes the feature fusion layer using a small number of real labeled samples, and prevents overfitting using a cosine annealing learning rate strategy. The final model simultaneously outputs three-scale feature maps and corresponding defect confidence heatmaps.

[0025] like Figure 1As shown, the core of multi-scale feature map fusion lies in the dynamic weight allocation strategy. First, the channel saliency score is calculated independently for each scale feature map: the activation value matrix of the feature map in the channel dimension is extracted, and the activation variance of all spatial locations of each channel is calculated; the variance values ​​are input into an S-shaped growth function to convert them into a normalized score between zero and one. This function is designed to have a smooth transition in low-variance regions and a steep rise in high-variance regions, ensuring that weak defect signals are not drowned out. The channel score matrices of the three sets of feature maps constitute the initial weight tensor. Before weighted fusion, cross-scale feature calibration is performed: the high-level semantic feature map is upsampled to the same size as the low-level feature map through bilinear interpolation, and the mid-level feature map recovers its detail resolution through transposed convolution; the calibrated feature maps are concatenated along the channel dimension to form the fusion basis tensor. Adaptive weighting operations apply the weight tensor to the fusion basis tensor: channels with scores higher than a set threshold are exponentially enhanced, with the enhancement coefficient increasing quadratically with the score exceeding the threshold; channels with scores lower than the threshold are linearly compressed. The weighted tensor undergoes channel reduction via a 1x1 convolutional layer, compressing it to one-quarter of the original number of channels. Batch normalization and exponential linear unit activation functions are then applied. The resulting fused feature signal contains three complementary types of information: defect geometric details preserved by low-level features, morphological topology characterized by mid-level features, and causal semantic attributes encoded by high-level features. This signal is transmitted to the decision module via a standardized interface, accompanied by a head region defect probability distribution map as an auxiliary criterion.

[0026] Furthermore, the graph neural network of the decision-making module constructs a heterogeneous defect knowledge graph as the basis for classification. The graph pattern layer defines four types of entity nodes: defect type nodes include standard classifications such as scratches, bubbles, and foreign objects; process factor nodes are associated with parameters such as extrusion temperature, die pressure, and raw material moisture content; equipment status nodes mark indicators such as die wear and roller concentricity; and material attribute nodes record data such as substrate batch and additive ratio. Edge relationships are defined as physical cause chains, such as scratch nodes being connected to die wear nodes through the relationship "caused by die damage," and bubble nodes being linked to moisture content nodes through the relationship "affected by raw material moisture content." After receiving the fused feature signal, the classifier first maps it to a 128-dimensional node attribute vector through a fully connected layer. The graph reasoning process adopts a multi-head graph attention mechanism: calculating the similarity weight between the attribute vector and each defect type node, and introducing process constraints in the weight calculation—if the current production line process parameters exceed the preset safety range of the knowledge graph, the similarity weight of the corresponding defect type is increased; and aggregating the information of adjacent nodes according to the weight distribution updates the feature representation of the defect type node. The classification output phase executes two layers of reasoning: the first layer matches the most likely defect type and selects the type node with the highest attention weight; the second layer backtracks along the edges of the knowledge graph to retrieve the set of process factor nodes that have a direct causal relationship with the selected defect type. The output signal is a structured data packet containing the main defect type label, a list of candidate type confidence scores, key process factor names, and influence strength coefficients. The system has an online knowledge update interface; when process engineers confirm new defect causal relationships, the graph size is expanded in real time by adding nodes and edges.

[0027] like Figure 2 The diagram illustrates the method for identifying and classifying defects on the prism film surface according to the present invention. S1: By collaboratively controlling multi-angle light sources, original image signals of the prism film surface under different optical conditions are acquired. S2: The original image signals are received, and a background model is constructed based on prior knowledge of the periodic microstructure of the prism film. Background interference is separated using frequency domain filtering and reflection compensation algorithms to generate a pre-processed image signal with suppressed background. S3: The pre-processed image signal is received, and multi-scale defect morphological and texture features are simultaneously extracted using a small-sample optimized deep learning model. The feature maps of different scales are adaptively weighted and fused to form a fused feature signal. S4: The fused feature signal is received, and the physical morphological features of the defects are associated with a preset defect cause knowledge base through a classification network. A defect type and potential process-related factor classification signal are output. S5: The factor classification signal is received, and a feedback instruction containing defect type, location coordinates, and process optimization suggestions is generated in real time. This instruction is then transmitted to the production line control system to trigger dynamic process parameter adjustments.

[0028] Specifically, the process rule base of this module adopts a three-level tree-like index structure. The first-level index divides rule groups according to defect type, and each group contains a historical case library and an expert experience matrix: the case library records process adjustment schemes and effect evaluation data for similar defects; the experience matrix defines the sensitivity ranking of process parameters to defects. The second-level index is associated with the process factors output by the decision-making module, and each factor is bound to an adjustment strategy table: for example, the extrusion temperature factor corresponds to the temperature compensation formula, and the die pressure factor is associated with the pressure gradient adjustment curve. The third-level index integrates spatial location data and maps defect coordinates to specific equipment stations in the production line digital twin model. The feedback instruction generation process is as follows: after receiving the classification signal, the main defect type is analyzed, and the basic adjustment direction is obtained by searching the first-level index; the top three key factors are selected according to the influence intensity coefficient of the process factors, and detailed adjustment strategies are matched in the second-level index; density clustering is performed in the spatial distribution model in combination with the defect location coordinates; if the same equipment station has multiple occurrences of the same type of defect within a set time window, an enhanced compensation amount is added to the strategy corresponding to that station. The optimization suggestion generation unit utilizes a natural language template engine to convert structured data such as process parameter adjustments, equipment maintenance actions, and raw material inspection requirements into readable text, such as "Add a 5-degree Celsius compensation to the third temperature zone of the extruder, shortening the mold polishing cycle to 70 hours." The final feedback instruction includes a machine-readable control instruction set and a human-readable operation manual, synchronously pushed to the production line control terminal and quality management terminal via an Industrial Internet of Things (IIoT) protocol. The system incorporates a feedback effectiveness evaluation loop: after instruction execution, it collects defect rate change data for subsequent production batches and automatically updates the case effectiveness evaluation values ​​in the rule base.

[0029] Furthermore, the data buffering and load balancing mechanism between the processing module and the fusion module is configured to dynamically optimize the allocation of computing resources. After receiving the preprocessed image signal output from the processing module, the data buffer unit performs a block reconstruction operation: the entire image is divided into several square regions of equal size, with adjacent blocks having a 10% overlap rate to avoid truncation due to boundary defects; each block independently calculates a background complexity score, based on a weighted sum of three dimensions—the variance of grayscale values ​​within the block characterizes texture richness, the Fourier spectral entropy value reflects periodicity intensity, and the proportion of edge pixels indicates structural density. Based on the score results, the blocks are divided into three levels: high-complexity blocks (top 20%) retain their original resolution and are marked as critical processing areas; medium-complexity blocks (middle 60%) are downsampled to three-quarters of their original resolution; and low-complexity blocks (bottom 20%) are downsampled to half their original resolution. The resource scheduler allocates feature extraction resources based on the hierarchical results: high-complexity blocks are allocated double the computation threads and a full-parameter model is enabled; medium-complexity blocks use standard threads and a pruned model; and low-complexity blocks are processed using a lightweight model with a single thread. Processed blocks are reconstructed into complete feature maps after weighted fusion of overlapping regions, with bicubic interpolation used to eliminate seam artifacts during reconstruction. The load monitoring unit tracks the processing latency of each block in real time. If the average processing time of a high-complexity block exceeds the system cycle constraint, its block size is automatically reduced and the number of parallel processing units is increased. Simultaneously, a historical complexity database is established, and membrane regions that repeatedly exhibit high complexity are pre-marked as permanent critical areas, allowing subsequent detection to skip the scoring stage and allocate the highest priority resources.

[0030] In one embodiment of the present invention, when the decision module receives the classification signal transmitted by the fusion module, a cross-modal verification process is initiated synchronously. The verification trigger monitors the classification difference of the same physical location under different imaging modalities in real time: extracting the defect type confidence vectors corresponding to the three modalities of coaxial light, low-angle light, and oblique light, and calculating the Euclidean distance between any two vectors; if any two sets of distances exceed a set threshold, it is determined to be a classification conflict event. The conflict processing unit immediately retrieves the multimodal original image of the corresponding location from the original image data warehouse and restarts the local processing process: performing background suppression and feature extraction only on the area surrounding the conflict location to generate high-precision local fusion features; inputting these features into three independently trained classification sub-networks (each sub-network specializing in one imaging modality) to obtain modality-specific classification results. The arbitrator compares the confidence distribution output by the sub-networks: selecting the classification result corresponding to the highest confidence value as the final decision, while recording the suboptimal results of other modalities as alternative annotations. The incremental learning engine is automatically triggered after arbitration: it packages the multimodal original images of the conflict cases, local fusion features, and arbitration results into incremental samples; it performs data augmentation operations on these samples, including random rotation, brightness perturbation, and adding Gaussian noise; it updates the training set of the small-sample optimized model and starts fine-tuning training, freezing the weights of the feature extraction layers during training and adjusting only the parameters of the last two fully connected layers to quickly adapt to new feature patterns. The system maintains a conflict case knowledge graph, recording the conflict frequency under different combinations of materials and defect types. When the conflict rate of a specific combination continues to rise, it automatically generates model structure optimization suggestions and submits them to the operation and maintenance terminal.

[0031] The prism film surface defect identification and classification management system and method of the present invention suppresses reflective interference through multi-angle light source collaborative imaging, uses prism periodic prior knowledge for frequency domain filtering to separate the background, adopts a small sample deep learning model to fuse multi-scale defect features, associates defect physical morphology with a knowledge base of causes to achieve interpretable classification, and finally generates feedback instructions containing process optimization suggestions to dynamically control the production line, forming a "detection-analysis-control" closed loop.

[0032] Therefore, the prism film surface defect identification and classification management system and method of the present invention solves the problem of high missed detection rate and inaccurate classification of small surface defects caused by strong reflection of microstructure and interference of complex texture of prism film.

[0033] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. A prism film surface defect identification and classification management system, characterized in that, include: A multimodal imaging module, which acquires raw image signals of the prism film surface under different optical conditions by coordinating the control of multi-angle light sources; The processing module receives the original image signal, constructs a background model based on prior knowledge of the periodic microstructure of the prism film, separates background interference through frequency domain filtering and reflection compensation algorithm, and generates a preprocessed image signal after background suppression. The fusion module receives the preprocessed image signal, uses a small-sample optimized deep learning model to simultaneously extract multi-scale defect morphological features and texture features, and adaptively weights and fuses the feature maps of different scales to form a fused feature signal. The decision module receives the fused feature signal, associates the physical morphological features of the defect with a preset defect cause knowledge base through a classification network, and outputs a defect type and potential process-related factor classification signal. The feedback module receives the factor classification signal, generates a feedback instruction in real time containing the defect type, location coordinates and process optimization suggestions, and transmits the instruction to the production line control system to trigger dynamic process parameter adjustment.

2. The prism film surface defect identification and classification management system according to claim 1, characterized in that, When the multimodal imaging module coordinates the control of multi-angle light sources, it is specifically configured to dynamically adjust the combination of incident angles and polarization directions of each light source according to the preset spatial frequency distribution of the microstructure on the prism film surface. The combination of incident angles includes coaxial illumination perpendicular to the film surface, low-angle grazing illumination less than ten degrees, and oblique diffuse illumination between thirty and sixty degrees. By triggering image sensors under different illumination modes in a time-division manner, multiple sets of original image signals with complementary defect response characteristics are acquired to ensure the imaging integrity of high reflective areas and low-contrast defect areas.

3. The prism film surface defect identification and classification management system according to claim 1, characterized in that, When constructing the background model, the processing module specifically performs Fourier transform analysis on the spatial frequency of the periodic microstructure of the prism film to extract the fundamental frequency and harmonic components as the background template. The frequency domain filtering algorithm selectively attenuates the frequency domain energy corresponding to the background template by designing a band-stop filter, while retaining the abnormal frequency domain components caused by defects. The reflection compensation algorithm uses the principle of polarization differential imaging to extract and suppress the specular reflection component from mutually orthogonal polarization channel images. The final preprocessed image signal with improved signal-to-noise ratio contains detextured defect enhancement information.

4. The prism film surface defect identification and classification management system according to claim 1, characterized in that, The fusion module employs a few-shot optimized deep learning model. It transfers the basic feature extraction layer of the pre-trained convolutional neural network and freezes its underlying weights to retain general texture recognition capabilities. During the model fine-tuning stage, a synthetic defect sample generation mechanism is introduced. This mechanism simulates the optical scattering model of scratches, bubbles, and foreign object attachment defects based on the physical properties of a prism film. The generated virtual defect images are fused with real, defect-free backgrounds to expand the training samples. At the model output, a multi-branch feature pyramid structure is deployed to extract morphological feature maps and local binary pattern texture feature maps under different receptive fields.

5. The prism film surface defect identification and classification management system according to claim 1, characterized in that, When the fusion module performs adaptive weighted fusion, it dynamically allocates weight coefficients based on the defect saliency heatmap of feature maps at each scale. The saliency heatmap is generated by calculating the variance of activation values ​​between feature map channels, and high-variance channels are given higher weights to enhance the response to minor defects. The weighted multi-scale feature maps are then spliced ​​together by channels and reduced in dimensionality by one-to-one convolution to form a fused feature signal that combines local details and global semantics.

6. The prism film surface defect identification and classification management system according to claim 1, characterized in that, The classification network of the decision module adopts a graph neural network architecture, which maps the fused feature signals to node attributes. The preset defect cause knowledge base is constructed as a heterogeneous knowledge graph, whose nodes include defect types, process parameters, and material attribute entities. The edge relationships define the physical association rules between defects and causes. The classification process calculates the similarity between the fused feature nodes and the knowledge graph nodes through a graph attention mechanism, and matches the defect type node with the highest similarity and its associated process factor nodes to achieve interpretable mapping from physical morphological features to causal factors.

7. The prism film surface defect identification and classification management system according to claim 1, characterized in that, When the feedback module generates process optimization suggestions, it performs multi-level searches in a preset process adjustment rule library based on the defect type and process correlation factor classification signals. The primary search matches the process parameter sensitivity ranking table corresponding to the defect type, the secondary search filters the top three key process parameters based on the correlation factor strength, and finally combines the clustering results of the defect location coordinates in the production line spatial distribution model to generate parameter tuning instructions for specific equipment.

8. The prism film surface defect identification and classification management system according to claim 1, characterized in that, The processing module and the fusion module are provided with a data buffer interface, which is configured to divide and reassemble the preprocessed image signal into blocks; divide the entire image into overlapping local region blocks, allocate different feature extraction computing resources according to the background complexity score of each region block, enable full-resolution processing for high-complexity regions and downsample processing for low-complexity regions, so as to achieve balanced computing load and real-time performance.

9. The prism film surface defect identification and classification management system according to claim 1, characterized in that, The decision module initiates an online verification mechanism after outputting the classification signal: when there is a conflict in the classification results of the same location under different imaging modalities, the reprocessing process of the multimodal original image signal is triggered. By comparing the confidence distribution of defect features under each modality, the classification result corresponding to the modality with the highest confidence is selected as the final output, and conflict cases are automatically added to the incremental training set of the small sample optimization model.

10. The intelligent suppression management method for the prism film surface defect identification and classification management system according to claims 1-9, comprising: S1: By coordinating the control of multi-angle light sources, the original image signals of the prism film surface under different optical conditions are acquired; S2: Receive the original image signal, construct a background model based on prior knowledge of the periodic microstructure of the prism film, separate background interference through frequency domain filtering and reflection compensation algorithm, and generate a preprocessed image signal after background suppression; S3: Receive the preprocessed image signal, use a small-sample optimized deep learning model to simultaneously extract multi-scale defect morphology features and texture features, and adaptively weight and fuse feature maps of different scales to form a fused feature signal; S4: Receive the fused feature signal, associate the physical morphological features of the defect with the preset defect cause knowledge base through the classification network, and output the defect type and potential process correlation factor classification signal. S5: Receive the factor classification signal, generate a feedback instruction in real time containing defect type, location coordinates and process optimization suggestions, and transmit the instruction to the production line control system to trigger dynamic process parameter adjustment.

Images