Diffusion plate quality monitoring system and method based on image analysis

Abstract

The invention relates to a diffusion plate quality monitoring system and method based on image analysis, and the system comprises a light source control module which generates a plurality of paths of driving signals and transmits the driving signals to a multi-mode light source device; the image acquisition module is used for receiving the multiple paths of driving signals, acquiring an original image of the diffusion plate in each illumination mode and generating a time sequence image sequence; the fusion calculation module is used for carrying out registration and feature extraction on the time sequence image sequence to generate a multi-channel fusion feature map; the defect analysis module is used for generating a defect position mask according to the gradient change and the texture features of the optical parameter distribution diagram and forming a quality analysis report containing defect types and coordinates; and the result output module is used for converting the quality analysis report into a visual detection result. According to the diffusion plate quality monitoring system and method based on image analysis, the problems of optical parameter quantitative detection and weak defect identification of the diffusion plate can be solved.

Description

Technical Field

[0001] This invention relates to the field of optical diffuser plate quality inspection technology, specifically to a diffuser plate quality monitoring system and method based on image analysis. Background Technology

[0002] As a core component of the backlight module, the diffuser plate's surface scratches, foreign objects, and other appearance defects, as well as its optical properties such as transmittance and haze, directly affect display uniformity. Traditional manual visual inspection is inefficient and cannot quantify optical parameters; contact integrating sphere measurement can only obtain an overall average value and damages the workpiece. Existing machine vision systems mostly rely on single-light source imaging, which is insufficient for detecting subtle scratches and foreign objects, and is particularly difficult to capture microscopic optical non-uniformities that cause bright and dark spots. Existing patents use ring light sources to enhance the contrast of surface defects, but do not solve the problem of detecting local anomalies in transmittance; the literature "Online Detection Technology of Optical Materials" proposes a transmission image analysis method; however, the internal scattering characteristics of the diffuser plate cause the image grayscale to have a non-linear relationship with the true optical parameters, resulting in significant errors in traditional calibration methods.

[0003] Existing patents attempt to identify foreign materials using multispectral imaging, but they lack a spatial distribution model of optical parameters, making it impossible to locate micron-level bright and dark spot defects. Furthermore, fluctuations in production line ambient lighting lead to inaccurate detection thresholds, and existing systems lack adaptability to novel defects. Especially for high-precision diffuser plates, stringent standards must be met simultaneously: a visual defect detection rate greater than 99.5% and local optical parameter fluctuations less than 2%. Existing technologies struggle to achieve this goal due to the lack of optical parameter inversion mechanisms and methods for analyzing the correlation of complex defects. Therefore, a detection solution integrating multimodal imaging, optical parameter spatial mapping, and dynamic threshold decision-making is urgently needed. Summary of the Invention

[0004] In view of the shortcomings of the prior art, the purpose of this invention is to provide an image analysis-based diffusion plate quality monitoring system and method to solve the problems of quantitative detection of optical parameters and identification of weak defects in diffusion plates. This invention acquires transmission and surface scattering image sequences by alternately illuminating the diffusion plate with multimodal light sources; it uses registration fusion and deep learning models to convert image features into spatially resolved transmittance / haze distribution maps; it dynamically generates defect masks based on the distribution map gradient and texture, and combines adaptive thresholds to identify both appearance and optical composite defects; finally, it outputs a visual report and sorting instructions, achieving non-contact online accurate detection.

[0005] This invention provides an image analysis-based diffusion plate quality monitoring system, comprising: The light source control module generates multiple driving signals and transmits them to the multi-mode light source device, driving it to output beams of at least two different illumination modes to illuminate the diffuser plate under test. The image acquisition module receives multiple drive signals, acquires the original image of the diffuser plate in each illumination mode, and generates a time-series image sequence. The fusion computing module performs registration and feature extraction on time-series image sequences to generate multi-channel fusion feature maps. Based on a pre-trained regression model, the fusion feature maps are converted into optical parameter distribution maps, which include spatially resolved transmittance distribution maps and haze distribution maps. The defect analysis module generates a defect location mask based on the gradient changes and texture features of the optical parameter distribution map. Combined with preset threshold conditions, it identifies appearance defects and optical performance abnormal areas, and generates a quality analysis report containing defect type and coordinates. The results output module converts the quality analysis report into visual inspection results and generates control instructions to be transmitted to the sorting execution mechanism.

[0006] In one embodiment of the present invention, the multimodal light source device driven by the light source control module includes an independently controllable parallel backlight unit and a ring light source unit; the parallel backlight unit forms a uniform surface light source through a collimating lens group to illuminate the back of the diffuser plate, which is used to acquire internal structural features in the transmission imaging mode; the ring light source unit surrounds the lens of the image acquisition module at a specific tilt angle, and its light-emitting elements are distributed in concentric circles. By adjusting the brightness ratio of each ring, the scattering characteristics of surface micro-scratches and foreign objects are enhanced; the two illumination modes are activated alternately according to a preset timing sequence, and the light source control module sends a synchronization pulse signal to the image acquisition module, so that the image acquisition module triggers exposure after each illumination mode stabilizes.

[0007] In one embodiment of the present invention, the image acquisition module includes a multispectral imaging unit and a high-speed trigger interface; the multispectral imaging unit is equipped with a switchable filter wheel, and acquires dual-channel images in the visible light band and near-infrared band under the ring light source illumination mode, and identifies the material type by comparing the reflectivity differences of foreign objects or contaminated areas under different bands; the high-speed trigger interface receives the pulse rising edge signal from the light source control module, and captures transient transmission images in the global shutter mode in the parallel backlight mode, and acquires high-resolution surface images in the area scanning mode in the ring light source mode, and finally classifies and encapsulates them into a time-series image sequence according to the brightness mode.

[0008] In one embodiment of the present invention, the feature extraction operation performed by the fusion computing module includes the construction of a multi-scale convolutional feature pyramid: first, the temporal image sequence is registered by sub-pixel-level affine transformation to eliminate pixel shifts caused by mechanical vibration; then, the gray-scale distribution features of the transmission image and the texture direction features of the surface image are extracted by parallel convolution paths, and the two types of feature maps are stitched together according to the channel dimension; finally, the spatial attention mechanism is used to weightedly fuse and stitch the features to generate a multi-channel fusion feature map that highlights the defect-sensitive area.

[0009] In one embodiment of the present invention, the pre-trained regression model is a dual-path deep neural network structure; the first path captures the macroscopic gradual change law of the optical parameter distribution map through a dilated convolutional layer, and the second path extracts microscopic local anomaly features through a residual connection structure; the outputs of the two paths are superimposed in the decoder stage, and spatially resolved transmittance distribution map and haze distribution map are reconstructed through a deconvolutional layer, wherein the transmittance distribution map maps the degree of light flux attenuation with gray values, and the haze distribution map characterizes the scattering uniformity with gradient magnitude.

[0010] In one embodiment of the present invention, the method for generating a defect location mask by the defect analysis module includes: performing a variable-direction Sobel operator convolution on the transmittance distribution map to detect transmittance abrupt change boundaries to form brightness abnormal regions; calculating local entropy values ​​on the haze distribution map to mark texture disordered regions as potential surface damage; superimposing the two types of regions and filling the gaps using morphological closing operations; and then aggregating adjacent pixels using a region growing algorithm to generate a connected defect location mask.

[0011] In one embodiment of the present invention, the threshold conditions for identifying appearance defects and optical performance abnormality areas by the defect analysis module include a dynamic adaptive mechanism: for scratch-type defects, the minimum aspect ratio and edge sharpness threshold of the mask area are dynamically adjusted according to the substrate type of the diffuser plate; for bright and dark spot defects, a warning line for the local deviation coefficient is set according to the overall standard deviation of the transmittance distribution map; when a certain area triggers both the appearance defect threshold and the optical performance abnormality threshold, it is marked as a composite defect and given the highest priority.

[0012] In one embodiment of the present invention, the visualization detection results of the result output module include a layered overlay display function: the original image is used as the bottom background, a semi-transparently rendered optical parameter distribution map is overlaid in the middle layer, and the mask boundary of the defect location is highlighted in the top layer; different color coding is used for different defect types, scratches are displayed as red polygonal boxes, bright spots are displayed as yellow contour lines, and dark spots are displayed as blue filled areas, while generating a machine-readable log containing defect coordinates and type codes.

[0013] In one embodiment of the present invention, the logic for generating control instructions by the result output module includes a sorting strategy mapping table: when the quality analysis report detects any optical performance abnormality in the central area of ​​the diffuser plate, a highest-level discard instruction is generated; when only the edge area has slight scratches and the light transmittance attenuation is lower than a set ratio, a rework instruction is generated and the coordinate range to be polished is marked; a qualified instruction is generated only when all defect types are acceptable minor foreign objects and have not triggered the optical performance threshold.

[0014] This invention also includes a diffusion plate quality monitoring method based on image analysis, comprising: S1: Generate multiple driving signals and transmit them to the multi-mode light source device to drive it to output beams of at least two different illumination modes to illuminate the diffuser plate under test; S2: Receives multiple drive signals, acquires the original image of the diffuser in each lighting mode, and generates a time-series image sequence; S3: Register and extract features from time-series image sequences to generate multi-channel fusion feature maps. Based on a pre-trained regression model, convert the fusion feature maps into optical parameter distribution maps, which include spatially resolved transmittance distribution maps and haze distribution maps. S4: Generate a defect location mask based on the gradient changes and texture features of the optical parameter distribution map, and identify appearance defects and optical performance abnormal areas by combining preset threshold conditions, and generate a quality analysis report containing defect type and coordinates. S5: Convert the quality analysis report into a visual inspection result and generate control instructions to be transmitted to the sorting execution mechanism.

[0015] This invention provides a diffusion plate quality monitoring system and method based on image analysis. The system acquires transmission and surface scattering image sequences by alternately illuminating the diffusion plate with multimodal light sources. It uses registration fusion and deep learning models to convert image features into spatially resolved transmittance / haze distribution maps. Based on the distribution map gradient and texture, a defect mask is dynamically generated, and combined with adaptive thresholds, appearance and optical composite defects are identified. Finally, a visual report and sorting instructions are output, realizing non-contact online accurate detection. Attached Figure Description

[0016] 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.

[0017] Figure 1 This is a system architecture diagram of a diffusion plate quality monitoring system based on image analysis; Figure 2 A schematic diagram illustrating the workflow of an image analysis-based diffusion plate quality monitoring system; Figure 3 This is a flowchart of a diffusion plate quality monitoring method based on image analysis. Detailed Implementation

[0018] 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.

[0019] 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.

[0020] 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.

[0021] Please see Figure 1-3 The image shows a diffusion plate quality monitoring system and method based on image analysis according to the present invention. The diffusion plate quality monitoring system based on image analysis of the present invention includes a light source control module, which generates multiple driving signals and transmits them to a multi-modal light source device to drive it to output beams of at least two different illumination modes to illuminate the diffusion plate under test; an image acquisition module, which receives the multiple driving signals, acquires the original image of the diffusion plate under each illumination mode, and generates a time-series image sequence; a fusion calculation module, which performs registration and feature extraction on the time-series image sequence to generate a multi-channel fusion feature map, and converts the fusion feature map into an optical parameter distribution map based on a pre-trained regression model, the optical parameter distribution map including a spatially resolved transmittance distribution map and a haze distribution map; a defect analysis module, which generates a defect location mask based on the gradient changes and texture features of the optical parameter distribution map, identifies appearance defects and optical performance abnormalities based on preset threshold conditions, and generates a quality analysis report containing defect type and coordinates; and a result output module, which converts the quality analysis report into a visualized detection result and generates control commands to be transmitted to the sorting execution mechanism.

[0022] like Figure 1As shown, an image analysis-based diffuser quality monitoring system achieves precise quantitative detection of diffuser optical performance and appearance defects through the collaborative operation of five core modules. The light source control module, as the system's starting point, generates multiple drive signals with phase differences, which are sent to the power controller of the multi-mode light source device via an electrically isolated transmission channel. This drive signal includes pulse width modulation waveforms and current amplitude parameters, allowing independent control of at least two illumination modes: the first mode activates the parallel backlight unit to generate a collimated beam, which is transmitted through the back of the diffuser to form a uniform light field; the second mode triggers the ring light source unit to project scattered light at an angle of 15 to 75 degrees, enhancing the optical response of the surface microstructure. The image acquisition module receives the synchronous trigger edge of the drive signal. In parallel backlight mode, a global shutter is used to capture the transmitted image, with the exposure time strictly matched to the light source pulse width to avoid motion blur; in ring light source mode, a rolling shutter is used to acquire high-resolution surface images through area scanning. All raw images are classified and encapsulated into a time-stamped image sequence according to their brightness mode. Images within the sequence are synchronized using hardware to ensure sub-millisecond alignment accuracy. After receiving the image sequence, the fusion computing module first performs sub-pixel-level affine transformation registration and uses a feature point matching algorithm to compensate for pixel shifts caused by mechanical vibration. Then, it initiates parallel processing channels: Channel 1 extracts the grayscale histogram features and frequency domain energy distribution of the transmission image, while Channel 2 analyzes the local binary pattern texture features and directional gradient histogram of the surface image. The dual-channel feature maps are then spatially pyramidal pooled and concatenated along the channel dimension to form a multi-channel fused feature map. This feature map is input into a pre-trained regression model—based on a deep convolutional neural network architecture. Its encoder contains twelve convolutional kernels to extract abstract features, and the decoder reconstructs the spatial mapping relationship through deconvolutional layers. Finally, it outputs a two-channel optical parameter distribution map: the first channel maps the relative transmittance of each region of the diffuser with eight-bit grayscale values, and the second channel encodes the non-uniformity of the haze distribution with gradient magnitude. The defect analysis module performs multi-scale analysis on the optical parameter distribution map: It applies an adjustable Sobel operator convolution kernel to the transmittance distribution map to detect areas of abrupt brightness changes and marks them as candidate bright or dark spots; it calculates the information entropy value within a sliding window on the haze distribution map to identify texture anomaly areas as suspected scratches or foreign objects; after superimposing the two types of regions, it performs a morphological closing operation to bridge the fracture boundaries, and then generates a connected defect location mask using a region growing algorithm. This mask is dynamically compared with preset threshold conditions: for bright spot defects, a deviation coefficient threshold is set based on the local standard deviation of the transmittance distribution map; for scratch defects, the severity level is determined based on the aspect ratio and edge sharpness of the mask area; finally, a quality analysis report is generated, including defect type classification codes, centroid coordinates, and area percentage. The results output module converts the report into a visualization interface, overlaying a semi-transparent rendered optical parameter heatmap on the original image and marking defect boundaries with red polygon boxes; simultaneously, it generates machine instruction codes to drive the sorting mechanism to perform discard, rework, or release actions based on the defect location and type.The entire system achieves 100-megabit-level data transmission between modules through a fiber optic network, ensuring that the online detection cycle is controlled within 500 milliseconds.

[0023] Furthermore, the multimodal light source device driven by the light source control module includes physically separated parallel backlight units and ring light source units. The parallel backlight unit consists of a densely arranged matrix of light-emitting diodes, each diode equipped with an independent reflector and Fresnel lens, forming a collimated beam with a divergence angle of less than three degrees through three optical designs; after passing through a frosted glass diffuser, the beam forms a surface light source with an illuminance uniformity of over 95% on the back of the diffuser plate. This unit is mounted on an air-floating translation stage, and the distance between the light source and the diffuser plate is adjusted by a servo motor to adapt to workpieces of different thicknesses. The ring light source unit adopts a four-ring concentric structure: the inner ring is arranged with blue light-emitting chips with a wavelength of 450 nanometers, the middle ring with green light-emitting chips with a wavelength of 530 nanometers, and the outer ring with red light-emitting chips with a wavelength of 620 nanometers. Each ring contains 120 independently dimmable light-emitting units; all chips are tilted at a 70-degree angle toward the center of the diffuser plate, and the light spot is softened by a quartz glass diffuser plate. The light source control module incorporates a timing logic controller: at the beginning of the detection cycle, it sends a zero-phase drive signal to the parallel backlight unit, maintaining a stable output for 50 milliseconds after activation; after the parallel backlight is turned off, it sends a 120-degree phase drive signal to the ring light source unit after a two-millisecond delay, illuminating the blue, green, and red light rings in a time-division manner. The brightness ratio of each ring is preset to 40% blue, 30% green, and 30% red, based on the diffuser material. A key innovation lies in the synchronization mechanism: when the parallel backlight reaches its rated illuminance, the light source control module sends a trigger pulse with a rising edge width of five nanoseconds to the image acquisition module; the switching intervals between the ring light sources generate falling edge pulses, ensuring the image acquisition module accurately captures each illumination state. Furthermore, the unit incorporates a temperature sensor to monitor the junction temperature of the LEDs in real time. If the detected temperature rise exceeds the safety threshold, a cooling cycle is automatically inserted, and brightness attenuation is dynamically compensated.

[0024] Specifically, the image acquisition module consists of a multispectral imaging unit and a high-speed trigger interface. The core of the multispectral imaging unit is a 20-megapixel scientific-grade image sensor, equipped with a double-layer coated protective window to reduce glare. A motorized filter wheel is mounted at the front of the lens, containing four filter slots: the first slot is a full-pass filter for parallel backlight mode; the second slot is a bandpass filter with a center wavelength of 450 nm and a bandwidth of 10 nm; the third slot is a center wavelength of 650 nm and a bandwidth of 20 nm; and the fourth slot is a near-infrared filter with a transmission range of 700 to 1000 nm. During the ring light illumination phase, the filter wheel switches time-division according to the control protocol: first, the full-pass filter is activated to acquire visible light composite images; then, it switches to a blue bandpass filter to match the inner ring blue light illumination; finally, the near-infrared filter is activated to detect the absorption characteristics of organic pollutants. The high-speed trigger interface employs fiber optic receiving technology to convert the pulse signal from the light source control module into an electrical trigger signal: upon receiving a parallel backlight trigger pulse, it immediately activates the global shutter exposure mode, with the exposure time programmable from one microsecond to one hundred milliseconds; upon receiving a ring light source trigger pulse, it switches to area scanning mode, dividing the sensor into sixteen partitions for parallel reading. The imaging unit is equipped with an active cooling system to maintain the sensor temperature at -10 degrees Celsius, significantly reducing dark current noise; the optical path consists of six sets of aspherical mirrors, achieving a color difference of less than two pixels in the 400-1000 nanometer wavelength range. The process of generating a time-series sequence from the original image includes three preprocessing steps: first, applying non-uniformity correction based on a dark field reference to eliminate pixel response differences; second, executing a bad pixel compensation algorithm to replace faulty pixels by weighting adjacent pixels; and finally, using a bilateral filter for noise reduction to suppress Gaussian noise while preserving edge features. All images are appended with metadata including illumination mode encoding, filter position, exposure parameters, and timestamps, and are packaged and transmitted to the fusion computing module in millisecond-level time windows.

[0025] In one embodiment of the present invention, the multi-scale convolutional feature pyramid construction process executed by the fusion computing module includes three processing stages. The first stage performs sub-pixel-level affine transformation registration: a robust feature algorithm is used to extract key points from the temporal image sequence, the optimal transformation matrix is ​​calculated using a random sampling consensus algorithm, and sub-pixel interpolation correction is applied to correct pixel-level offsets caused by mechanical vibration, ensuring that the spatial correspondence error between the transmission image and the surface image is less than 0.3 pixels. The second stage initiates parallel convolutional paths: Path one processes the transmission image simultaneously using three convolutional kernels of different scales—a 5x5 convolutional kernel extracts macroscopic grayscale distribution features, a 3x3 convolutional kernel captures mesoscale transmissive anomaly regions, and a 1x1 convolutional kernel preserves microscopic details; Path two processes the surface image, applying directionally separable convolutional layers to separate horizontal and vertical texture features, and then quantifying local scattering characteristics using gradient direction histograms. The output feature maps of the two paths are batch normalized and then stitched together along the channel dimension to form a primary fusion feature map. The third stage implements a spatial attention mechanism: First, global average pooling is performed on the primary fusion feature map to generate channel weight vectors. Then, a fully connected layer learns the defect sensitivity coefficients of each channel. After multiplying the weight vectors with the primary feature map channel by channel, bilinear interpolation upsampling is used to restore the spatial resolution, ultimately generating a 64-channel multi-channel fusion feature map. This feature map is superimposed with the residuals of the primary feature map through a skip connection mechanism to ensure that the features of small defects are not diluted by the deep network. The top layer of the feature pyramid outputs high-dimensional abstract features downsampled by eight times, the bottom layer retains the detailed features of the original resolution, and the middle layer achieves adaptive matching of feature scale through max pooling and transposed convolution, forming a multi-scale detection capability covering defects from millimeters to micrometers.

[0026] like Figure 2As shown, the pre-trained regression model employs a dual-path deep neural network architecture to achieve high-precision reconstruction of the optical parameter distribution map. The first path is the macroscopic gradient capture path: after inputting the multi-channel fused feature map, it sequentially passes through four sets of dilated convolutional modules. Each module contains two 3x3 dilated convolutional layers with dilation coefficients of 2 and 4, respectively, combined with batch normalization and exponential linear unit activation functions. By progressively increasing the dilation coefficient, the receptive field is gradually expanded to one-third of the full image size, specifically for extracting the spatial gradient trends of transmittance and haze. The second path is the microscopic anomaly extraction path: eight residual blocks are constructed using a residual connection structure. Each residual block contains two 1x1 convolutional layers and an identity mapping short-circuit connection, focusing on capturing abrupt changes in optical parameters in local regions. The outputs of the two paths are fused in the decoder stage: first, the macroscopic path output is reduced to 32 channels by 1x1 convolution, and the microscopic path output is increased to the same number of channels by 3x3 convolution; then, element-wise addition is performed, and the fused feature map is input into the reconstruction decoder. The reconstruction decoder consists of five upsampling modules: each module first doubles the feature map size through bilinear interpolation, then refines the features through two 3x3 convolutional layers; after the final upsampling stage, a 1x1 convolutional layer generates a dual-channel output. The first channel is normalized to a transmittance distribution map in the zero-to-one range using a Sigmoid activation function, where the gray value of each pixel represents the relative transmittance at that location; the second channel outputs a haze distribution map using a ReLU activation function, where the pixel values ​​represent the magnitude of the local scattering gradient. Model training employs a multi-task loss function: the transmittance channel uses a weighted combination of structural similarity index and mean squared error, while the haze channel uses a smoothed L1 loss function; simultaneously, a gradient difference regularization term is added to constrain the physical continuity of the distribution map.

[0027] Furthermore, the specific process of generating a defect location mask in the defect analysis module includes four calculation steps. Step 1: Transmittance anomaly detection: Perform directional variable Sobel operator convolution on the transmittance distribution map, and calculate gradient components in four directions: 0°, 45°, 90°, and 135°. Take the maximum gradient value of each pixel to form an initial edge intensity map, and refine the edge lines by suppressing non-maximum values. Perform adaptive threshold binarization on the refined edge map, and mark pixels with gradient values ​​greater than twice the standard deviation of the local window mean as brightness anomaly boundaries. Step 2: Texture disorder region marking: For the haze distribution map, calculate the Shannon information entropy value within a 9x9 sliding window. The entropy value calculation formula is the negative logarithmic weighted sum of local probability distributions. When the window entropy value exceeds 30% of the overall image entropy value, it is determined that there is a texture anomaly in the region. Perform morphological opening operation on the abnormal region to remove noise points and generate a potential surface damage binary map. Step 3: Defect Region Fusion. A logical OR operation is performed on the brightness anomaly boundary map and the texture anomaly map to merge the two types of suspicious regions. Morphological closing operations are then applied to the merged result, using a 5x5 circular structuring element for first dilation and then erosion to bridge fracture boundaries and fill micro-holes. Step 4: Connected Region Aggregation. Using the foreground pixels in the binary image output from Step 3 as seed points, an eight-neighborhood region growing algorithm is initiated. The growth threshold is set to ensure that the transmittance difference between adjacent pixels is less than 5% and the consistency of the haze gradient direction is greater than 0.7. After growth, the minimum bounding rectangle is calculated for each connected region. When the rectangle area is greater than twenty pixels, the output is the final defect location mask. This mask is a single-channel binary image, with white areas representing identified potential defect regions.

[0028] like Figure 3 The diagram illustrates a diffusion plate quality monitoring method based on image analysis according to the present invention. S1: Generate multiple driving signals and transmit them to a multimodal light source device, driving it to output beams of at least two different illumination modes to illuminate the diffusion plate under test; S2: Receive the multiple driving signals, acquire the original image of the diffusion plate under each illumination mode, and generate a time-series image sequence; S3: Register and extract features from the time-series image sequence to generate a multi-channel fusion feature map. Based on a pre-trained regression model, convert the fusion feature map into an optical parameter distribution map, which includes a spatially resolved transmittance distribution map and a haze distribution map; S4: Generate a defect location mask based on the gradient changes and texture features of the optical parameter distribution map, and identify appearance defects and optical performance abnormalities by combining preset threshold conditions, forming a quality analysis report containing defect types and coordinates; S5: Convert the quality analysis report into a visualized detection result and generate control commands to transmit to the sorting execution mechanism.

[0029] Specifically, the adaptive threshold mechanism of the defect analysis module includes dynamic discrimination logic for three types of defects. For scratch-type defects: First, the geometric features of each connected region are extracted from the defect location mask, including the aspect ratio of the minimum bounding rectangle, the gradient direction consistency of edge pixels, and the ratio of region area to perimeter. When the aspect ratio is greater than 3 and the gradient direction consistency is greater than 0.6, it is initially judged as a linear scratch. Further, the grayscale profile curve along the main direction of the scratch is calculated. If the curve exhibits the typical characteristics of steep gradients on both sides and a depression in the middle, it is confirmed as a scratch defect. Finally, based on the ratio of the depression depth to the substrate transmittance, it is classified into three levels: minor scratch, moderate scratch, and severe scratch. For the determination of bright and dark spot defects: The corresponding area of ​​the mask at the defect location is cropped from the transmittance distribution map. The difference between the average transmittance of this area and the average transmittance of the surrounding annular reference area with a diameter of three times is calculated. When the difference is greater than 1.5 times the standard deviation of the overall image transmittance, it is determined as a bright spot; when it is less than -1.5 times, it is determined as a dark spot. The severity of bright spots is graded according to the proportion of local transmittance exceeding the standard, and the severity of dark spots is graded according to the proportion of transmittance loss. For the association rules of composite defects: When a region simultaneously meets the geometric characteristics of a scratch and the optical characteristics of a bright / dark spot, a composite analysis process is initiated—the transmittance change curve along the main direction of the scratch is calculated. If the curve shows a sudden increase or decrease in transmittance at the center of the scratch, it is determined as an optically damaged scratch; if the transmittance curve is consistent with the background area, it is classified as a simple surface scratch. All determination results are dynamically associated with the diffuser plate type: For polycarbonate substrates, the scratch depth threshold is automatically reduced by 15%, and for acrylic substrates, the bright spot detection sensitivity is increased by 20%. When a new undefined defect pattern is detected, the sample feature vector is automatically stored and an incremental learning flag is triggered.

[0030] Furthermore, the visualization of the detection results in the output module enables layered overlay display, and its rendering process includes a 3D data synthesis and interactive control mechanism. First, a display coordinate system is established: the original acquired image is used as the bottom reference layer, scaled to the display resolution using bilinear interpolation; the intermediate layer loads optical parameter distribution map data, with the transmittance distribution map mapped using a rainbow color scheme (deep blue for low transmittance areas, bright red for high transmittance areas), and the haze distribution map mapped using a grayscale gradient (light gray for low haze, dark gray for high haze). Both are then overlaid onto the bottom image with 40% transparency using an alpha channel mixer. The top-level defect annotation layer is dynamically generated: the defect location mask extracts polygon vertices using an edge tracking algorithm; for scratch-type defects, a red Bezier curve is used to draw smooth boundaries, with the line width positively correlated with the scratch depth; for bright spot defects, contour lines are extracted from the transmittance distribution map, and closed areas exceeding a threshold of 15% are filled with a yellow semi-transparent overlay layer; for dark spot defects, local transmittance valley areas are calculated and marked with blue diagonal lines. Machine-readable logs are generated synchronously: Each defect entry contains a 16-byte header identifier, followed by the defect type code (a single hexadecimal digit identifying scratches, bright spots, dark spots, etc.), centroid coordinates (four-byte floating-point numbers storing X and Y values), circumscribed rectangle dimensions (four bytes storing length and width), area percentage (two bytes storing percentage), and optical parameter offset (four bytes storing transmittance and haze deviation). Log files are named according to the inspection timestamp and transmitted to the Manufacturing Execution System database via Industrial Ethernet. The visual interface is equipped with interactive controls: users can drag the transparency slider to adjust the development intensity of the optical parameter layer, check checkboxes to switch the display status of specific defect types, and double-click the defect area to bring up a detailed parameter pop-up window; all user operation events are recorded in the audit log for quality traceability.

[0031] This invention discloses a diffusion plate quality monitoring system and method based on image analysis. The system acquires transmission and surface scattering image sequences by alternately illuminating the diffusion plate with multimodal light sources. It then uses registration fusion and deep learning models to convert image features into spatially resolved transmittance / haze distribution maps. Based on the distribution map gradient and texture, a defect mask is dynamically generated, and combined with adaptive thresholds, appearance and optical composite defects are identified. Finally, a visual report and sorting instructions are output, achieving non-contact online accurate detection.

[0032] Therefore, the present invention provides a diffusion plate quality monitoring system and method based on image analysis, which solves the problems of quantitative detection of optical parameters and identification of weak defects in diffusion plates.

[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 diffusion plate quality monitoring system based on image analysis, characterized in that, include: A light source control module generates multiple driving signals and transmits them to a multi-mode light source device to drive it to output beams of at least two different illumination modes to illuminate the diffuser plate under test. An image acquisition module receives the multiple driving signals, acquires the original image of the diffuser plate in each illumination mode, and generates a time-series image sequence. The fusion computing module performs registration and feature extraction on the temporal image sequence to generate a multi-channel fusion feature map. Based on a pre-trained regression model, the fusion feature map is converted into an optical parameter distribution map, which includes a spatially resolved transmittance distribution map and a haze distribution map. The defect analysis module generates a defect location mask based on the gradient changes and texture features of the optical parameter distribution map, and identifies appearance defects and optical performance abnormal areas by combining preset threshold conditions, thus generating a quality analysis report containing defect type and coordinates. The result output module converts the quality analysis report into visual inspection results and generates control instructions that are transmitted to the sorting execution mechanism.

2. The diffusion plate quality monitoring system based on image analysis according to claim 1, characterized in that, The multimodal light source device driven by the light source control module includes an independently controllable parallel backlight unit and a ring light source unit. The parallel backlight unit forms a uniform surface light source through a collimating lens group to illuminate the back of the diffuser plate, which is used to acquire internal structural features in the transmission imaging mode. The ring light source unit surrounds the lens of the image acquisition module at a specific tilt angle, and its light-emitting elements are distributed in concentric circles. By adjusting the brightness ratio of each ring, the scattering characteristics of surface micro-scratches and foreign objects are enhanced. The two illumination modes are activated alternately according to a preset timing sequence. The light source control module sends a synchronization pulse signal to the image acquisition module, so that the image acquisition module triggers exposure after each illumination mode stabilizes.

3. The diffusion plate quality monitoring system based on image analysis according to claim 1, characterized in that, The image acquisition module includes a multispectral imaging unit and a high-speed trigger interface. The multispectral imaging unit is equipped with a switchable filter wheel, which acquires dual-channel images in the visible light and near-infrared bands under ring light illumination mode. The material type is identified by comparing the reflectivity differences of foreign objects or contaminated areas under different bands. The high-speed trigger interface receives the pulse rising edge signal from the light source control module. In parallel backlight mode, it uses a global shutter mode to capture transient transmission images. In ring light mode, it uses a region scanning mode to acquire high-resolution surface images. Finally, the images are classified and packaged into a time-series image sequence according to the illumination mode.

4. The diffusion plate quality monitoring system based on image analysis according to claim 1, characterized in that, The feature extraction operation performed by the fusion computing module includes the construction of a multi-scale convolutional feature pyramid: first, sub-pixel-level affine transformation registration is performed on the temporal image sequence to eliminate pixel shifts caused by mechanical vibration; then, the gray-level distribution features of the transmission image and the texture direction features of the surface image are extracted through parallel convolution paths, and the two types of feature maps are stitched together according to the channel dimension; finally, a spatial attention mechanism is used to weighted fuse and stitch the features to generate a multi-channel fusion feature map that highlights the defect-sensitive area.

5. The diffusion plate quality monitoring system based on image analysis according to claim 1, characterized in that, The pre-trained regression model is a dual-path deep neural network structure; the first path captures the macroscopic gradual change pattern of the optical parameter distribution map through a dilated convolutional layer, and the second path extracts microscopic local anomaly features through a residual connection structure; the outputs of the two paths are superimposed in the decoder stage, and spatially resolved transmittance distribution map and haze distribution map are reconstructed through a deconvolutional layer, wherein the transmittance distribution map maps the degree of light flux attenuation with gray values, and the haze distribution map characterizes the scattering uniformity with gradient magnitude.

6. The diffusion plate quality monitoring system based on image analysis according to claim 1, characterized in that, The method for generating a defect location mask by the defect analysis module includes: performing a variable-direction Sobel operator convolution on the transmittance distribution map to detect transmittance abrupt change boundaries and form brightness abnormal regions; calculating the local entropy value on the haze distribution map and marking texture disordered regions as potential surface damage; superimposing the two types of regions and filling the gaps with morphological closing operations, and then aggregating adjacent pixels through a region growing algorithm to generate a connected defect location mask.

7. The diffusion plate quality monitoring system based on image analysis according to claim 1, characterized in that, The defect analysis module identifies appearance defects and optical performance abnormality areas using threshold conditions with a dynamic adaptive mechanism: for scratch-like defects, the minimum aspect ratio and edge sharpness threshold of the mask area are dynamically adjusted according to the substrate type of the diffuser plate; for bright and dark spot defects, a warning line for the local deviation coefficient is set based on the overall standard deviation of the transmittance distribution map; when a certain area triggers both the appearance defect threshold and the optical performance abnormality threshold, it is marked as a composite defect and given the highest priority.

8. The diffusion plate quality monitoring system based on image analysis according to claim 1, characterized in that, The visualization of the results output module includes a layered overlay display function: the original image is used as the bottom background, a semi-transparent rendered optical parameter distribution map is overlaid in the middle layer, and the mask boundary of the defect location is highlighted in the top layer; different color codes are used for different defect types, scratches are displayed as red polygon boxes, bright spots are displayed as yellow contour lines, and dark spots are displayed as blue filled areas, while generating machine-readable logs containing defect coordinates and type codes.

9. The diffusion plate quality monitoring system based on image analysis according to claim 1, characterized in that, The logic for generating control instructions in the result output module includes a sorting strategy mapping table: when the quality analysis report detects any optical performance abnormality in the central area of ​​the diffuser plate, the highest-level discard instruction is generated; when only the edge area has slight scratches and the light transmittance attenuation is lower than the set ratio, a rework instruction is generated and the coordinate range to be polished is marked; only when all defect types are acceptable minor foreign objects and do not trigger the optical performance threshold, a qualified instruction is generated.

10. An intelligent suppression management method for a diffuser plate quality monitoring system based on image analysis as described in claims 1-9, comprising: S1: Generate multiple driving signals and transmit them to the multi-mode light source device to drive it to output beams of at least two different illumination modes to illuminate the diffuser plate under test; S2: Receive the multi-channel drive signal, acquire the original image of the diffuser plate in each lighting mode, and generate a time-series image sequence; S3: Register and extract features from the time-series image sequence to generate a multi-channel fusion feature map. Based on a pre-trained regression model, convert the fusion feature map into an optical parameter distribution map, which includes a spatially resolved transmittance distribution map and a haze distribution map. S4: Generate a defect location mask based on the gradient changes and texture features of the optical parameter distribution map, identify appearance defects and optical performance abnormal areas by combining preset threshold conditions, and form a quality analysis report containing defect type and coordinates. S5: Convert the quality analysis report into a visual inspection result and generate control instructions to transmit to the sorting execution mechanism.

Images