Ceramic appearance automatic detection method and system based on machine vision

Abstract

The invention relates to a ceramic appearance automatic detection method and system based on machine vision. The method comprises the following steps: arranging a multi-light-source combination for illumination, optimizing light source parameters, and dynamically adjusting light source intensity according to a ceramic material; ceramic images are collected, and meanwhile a timestamp and a light source type are marked for each frame of image; multi-light-source images are subjected to feature point matching, and weighted fusion is carried out to obtain a fused image; inputting the fused image into the semantic segmentation model, and outputting a defect category and a position; and the semantic segmentation model can automatically process the fused image and quickly output the detection result, so that compared with manual detection, the detection efficiency is greatly improved, and the requirements of large-scale ceramic production are met.

Description

Technical Field

[0001] This invention relates to the field of magnetic gradient measurement technology, and in particular to an automatic inspection method and system for ceramic appearance based on machine vision. Background Technology

[0002] Limitations of traditional ceramic appearance inspection methods: Manual inspection: For a long time, ceramic appearance inspection has mainly relied on manual visual observation. However, manual inspection has many drawbacks. On the one hand, manual inspection is inefficient and cannot meet the needs of large-scale production. On a ceramic production line, a large number of ceramic products may be produced every minute, and manual inspection cannot keep up with the production speed. On the other hand, manual inspection is easily affected by subjective factors. Different inspectors may have different judgment standards, and after working for a long time, inspectors are prone to fatigue, leading to a decrease in inspection accuracy and making it difficult to ensure the consistency of product quality.

[0003] Single-light source machine vision inspection: Early machine vision inspection technologies primarily used a single light source for illumination. Single-light source inspection of ceramic appearance has limitations. For different types of defects, such as cracks, dents, and spots, a single light source may not be able to clearly highlight all defect features. For example, some defects may be obscured by shadows under specific lighting angles, leading to missed detection. Furthermore, single-light source inspection is poorly adaptable to different ceramic materials, and cannot dynamically adjust the lighting effect according to changes in the ceramic material.

[0004] Existing patent application CN202510751182.4 discloses a visual inspection system and method for ceramic products, including a multi-angle illumination module, a multispectral three-dimensional imaging module, a feature separation and extraction module, a depth feature analysis module, a multimodal fusion decision module, an adaptive detection control module, and a data processing and storage module. This invention simultaneously acquires multi-band images and depth maps through the multispectral three-dimensional imaging module, accurately identifies defects from the feature and depth dimensions using the feature separation and extraction module and the depth feature analysis module respectively, and then integrates the information through the multimodal fusion decision module to output a reliable judgment. The adaptive detection control module adjusts its strategy based on the product ID to adapt to diverse needs. The modules work collaboratively to achieve high-precision, intelligent detection, effectively solving the problem of low accuracy in traditional detection methods, improving detection efficiency and accuracy, and providing strong quality assurance for ceramic production. However, it still has certain shortcomings.

[0005] Therefore, this invention proposes an automatic inspection method and system for ceramic appearance based on machine vision. Summary of the Invention

[0006] This invention addresses the technical problems existing in the prior art by providing a method and system for automatic inspection of ceramic appearance based on machine vision.

[0007] The technical solution of this invention to solve the above-mentioned technical problems is as follows: an automatic inspection method and system for ceramic appearance based on machine vision; wherein the method includes the following steps: S1: Set up a combination of multiple light sources for illumination, optimize the light source parameters, and dynamically adjust the light source intensity according to the ceramic material; S2: Acquire ceramic images and simultaneously mark each frame with a timestamp and light source type; S3: The multi-source images are matched by feature points and then weighted and fused to obtain a fused image; S4: Input the fused image into the semantic segmentation model, and output the defect category and location.

[0008] Furthermore, in the aforementioned automatic ceramic appearance inspection method based on machine vision, the multi-light source combination includes: a ring light source, a coaxial light source, and a low-angle linear light source.

[0009] Furthermore, in the aforementioned automatic ceramic appearance inspection method based on machine vision, the optimization of light source parameters further includes: Optimize the angles of multiple light sources separately: ring light source angle: 30-45°, low-angle light source angle: 10-20°; Multi-source timing control: Time-sharing triggering: Switching light sources at different times to capture images under different lighting conditions; Synchronous triggering: Multiple light sources work simultaneously, acquiring images from multiple light sources at the same time.

[0010] Furthermore, in the aforementioned automatic ceramic appearance inspection method based on machine vision, the step of dynamically adjusting the light source intensity according to the ceramic material includes: using PID feedback control to adjust the light source intensity in real time. ; in It is the error between the current image grayscale variance and the target variance. , , These are PID coefficients. It is a differential term.

[0011] Furthermore, the aforementioned automatic ceramic appearance detection method based on machine vision includes feature point matching for its multi-light source images, including: Align multi-source images to ensure accurate pixel correspondence during subsequent fusion, detect key points, and generate 128-dimensional feature descriptors for feature matching: ; in It is a feature descriptor of the reference image. It is the nearest neighbor feature descriptor of the image to be matched. It is the second nearest neighbor feature descriptor of the image to be matched. It is a threshold.

[0012] Furthermore, the aforementioned machine vision-based automatic ceramic appearance detection method, through weighted fusion, obtains a fused image that includes: adaptively fusing multi-source images to highlight defect features and suppress noise. ; in, This is the image under the k-th light source. Let n be the weight of the k-th light source, and n be the number of light sources. Weight calculation based on signal-to-noise ratio: , ; in, It is the standard deviation of grayscale in the defective area. It is the standard deviation of grayscale in the background area; Furthermore, in the aforementioned automatic ceramic appearance detection method based on machine vision, the semantic segmentation model includes: a backbone network, an ASPP module, and a decoder; Multi-scale features are extracted from the backbone network while dilated convolution is performed to avoid resolution loss due to downsampling. The ASPP module uses parallel multi-branch dilated convolution to capture multi-scale contextual information. The decoder fuses shallow and deep features.

[0013] On the other hand, a machine vision-based automatic ceramic appearance inspection system is provided, applied to the aforementioned machine vision-based automatic ceramic appearance inspection method, the system comprising: Lighting module: Employs a combination of various light sources to meet lighting needs from different angles and in different ways; Light source controller: Used to adjust the parameters of the light source, and dynamically adjusts the light source intensity according to the different ceramic materials to obtain the best lighting effect; Sensor and feedback device: Equipped with light sensors to monitor the light conditions of the lighting environment in real time and feed the data back to the light source controller; Image acquisition module: Acquires ceramic images and marks each frame with a timestamp and light source type; Fusion module: It matches multi-source images using feature points and performs weighted fusion to obtain a fused image; Output module: Input the fused image into the semantic segmentation model, and output the defect category and location.

[0014] The beneficial effects of this invention are: Improving Defect Display: A multi-source light source combination, including a ring light source, a coaxial light source, and a low-angle linear light source, is employed with optimized light source parameters to illuminate the ceramics from different angles and directions. The ring light source provides uniform ring illumination, helping to highlight the overall features of the ceramic surface; the coaxial light source reduces interference from reflected light, clearly displaying minute defects on the ceramic surface; and the low-angle linear light source emphasizes surface irregularities. Timing control through time-division and synchronous triggering allows for the acquisition of images under different lighting conditions, enriching image information and improving defect display. Simultaneously, the light source intensity is dynamically adjusted according to the ceramic material, using PID feedback control for real-time adjustment, ensuring the lighting effect adapts to different ceramic materials. Suitable lighting conditions are provided for ceramic products with varying surface gloss and color depth, ensuring defects are clearly displayed and improving the adaptability and accuracy of the inspection.

[0015] Multi-source images are aligned through feature point matching and then weighted and fused to obtain a fused image. The adaptive fusion method can rationally allocate weights based on the characteristics of different light source images, highlighting defect features while suppressing noise interference. A signal-to-noise ratio-based weight calculation method automatically adjusts the weights of each light source image according to the grayscale standard deviation of the defect and background regions, making the fused image more conducive to defect detection and recognition. The semantic segmentation model adopts a structure of a backbone network, an ASPP module, and a decoder. The backbone network extracts multi-scale features and performs dilated convolutions, avoiding downsampling loss of resolution and preserving more image detail information. The ASPP module uses parallel multi-branch dilated convolutions to capture multi-scale contextual information, enabling the model to better understand the semantic information in the image and accurately identify different types of defects. The decoder fuses shallow and deep features, combining high-resolution shallow features with strong semantic deep features, further improving the accuracy of defect detection and accurately outputting the category and location of defects. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating an automatic ceramic appearance inspection method based on machine vision. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0019] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0020] An automatic inspection method and system for ceramic appearance based on machine vision; the method includes the following steps: S1: Set up a combination of multiple light sources for illumination, optimize the light source parameters, and dynamically adjust the light source intensity according to the ceramic material; S2: Acquire ceramic images and simultaneously mark each frame with a timestamp and light source type; S3: The multi-source images are matched by feature points and then weighted and fused to obtain a fused image; S4: Input the fused image into the semantic segmentation model, and output the defect category and location.

[0021] The multi-light source combination includes: a ring light source, a coaxial light source, and a low-angle linear light source.

[0022] Specifically: Ring light sources, coaxial light sources, and low-angle linear light sources are used; different types of light sources illuminate the ceramics from different angles and in different ways. Ring light sources provide uniform circumferential illumination, coaxial light sources are suitable for detecting minute defects on smooth surfaces, and low-angle linear light sources can highlight surface irregularities. High-precision image acquisition equipment (such as industrial cameras) is used to acquire images of the ceramics. During the acquisition process, each frame of the image is marked with a timestamp and the light source type. The timestamp records the specific time of image acquisition, facilitating subsequent image sorting and tracing. The light source type information helps distinguish images acquired under different lighting conditions, providing a necessary basis for subsequent image fusion. Key points are detected in multi-light source images, and 128-dimensional feature descriptors are generated. These feature descriptors describe the image features around the key points. The distance relationship between feature descriptors in different images is compared to determine whether the feature points match. The multi-light source images are accurately aligned to ensure that pixels correspond one-to-one during subsequent fusion. Different weights are assigned to the images based on the signal-to-noise ratio of the images under different light sources. The standard deviation of gray levels in the defect area and the background area are calculated, and weights are determined based on these two values. Multi-light source images with different weights are fused to highlight defect features while suppressing the influence of noise, resulting in a fused image that integrates multiple illumination information. Finally, the fused image is input into a semantic segmentation model. After processing and analysis by the model, the final output is the category (such as scratches, cracks, protrusions, dents, etc.) and location (represented in the form of coordinates in the image), providing accurate information for the quality inspection of ceramics.

[0023] Specifically, the optimization of light source parameters also includes: The angles of multiple light sources are optimized separately. The angle of the ring light source is 30-45°. This angle range can make the light cover the ceramic surface evenly and avoid uneven lighting caused by excessively large or small angles. A low-angle light source (10-20°) can maximize the highlighting of surface undulations. Multi-source timing control: Time-sharing triggering: The light source is switched at different times to capture images under different lighting conditions; different light sources are triggered sequentially at different times, with only one light source working at a time and capturing the corresponding image. The advantage of this is that it avoids interference between different light sources, ensuring that each captured image is a clear image under a single light source.

[0024] Synchronous triggering: Multiple light sources operate simultaneously, acquiring images from all sources at once; multiple light sources work concurrently, acquiring images under their combined influence. This method can quickly obtain comprehensive information about ceramics under various lighting conditions, improving detection efficiency.

[0025] Specifically, the dynamic adjustment of light source intensity based on ceramic material includes: using PID feedback control to adjust the light source intensity in real time. ; in It is the error between the current image grayscale variance and the target variance. , , These are PID coefficients. It's the differential term. A PID feedback control algorithm is used to adjust the light source intensity in real time. First, the gray-level variance of the currently acquired image is calculated and compared with a pre-set target variance to obtain the error value. Then, the error is processed according to the PID coefficients and the differential term, dynamically adjusting the light source intensity to achieve the optimal gray-level distribution of the image, which is more conducive to subsequent defect identification.

[0026] Specifically, feature point matching is used to analyze multi-source images, including: Align multi-source images to ensure accurate pixel correspondence during subsequent fusion, detect key points, and generate 128-dimensional feature descriptors for feature matching: ; in It is a feature descriptor of the reference image. It is the nearest neighbor feature descriptor of the image to be matched. It is the second nearest neighbor feature descriptor of the image to be matched. It is a threshold.

[0027] Specifically, the weighted fusion process yields a fused image that includes: adaptive fusion of multi-source images to highlight defect features and suppress noise. ; in, This is the image under the k-th light source. Let n be the weight of the k-th light source, and n be the number of light sources. Weight calculation based on signal-to-noise ratio: , ; in, It is the standard deviation of grayscale in the defective area. It is the standard deviation of grayscale in the background area.

[0028] Specifically, the semantic segmentation model includes: a backbone network, an ASPP module, and a decoder; Multi-scale features are extracted from the backbone network while dilated convolution is performed to avoid resolution loss due to downsampling. The ASPP module uses parallel multi-branch dilated convolution to capture multi-scale contextual information. The decoder fuses shallow and deep features.

[0029] In this semantic segmentation model architecture (taking DeepLabV3+ as an example); Objective: To classify each pixel and output the defect category (e.g., crack, dent) and location (pixel-level mask). Backbone network ResNet-101 or MobileNetV2 can be used to extract multi-scale features.

[0030] Atrous convolution expands the receptive field and avoids resolution loss during downsampling: y[i]=\sum_{k}x[i+r\cdotk]\cdotw[k]; Where y[i] is the value at the i-th position of the output feature map. ASPP module (Atrous Spatial Pyramid Pooling); x is the input feature map, w[k] is the k-th weight of the convolution kernel, and r is the dilation rate; Parallel multi-branch dilated convolutions (with different dilation rates (r=6, 12, 18)) capture multi-scale contextual information: \text{ASPP}(x)=\text{Concat}\left[\text{Conv}_{r=6}(x),\text{Conv}_{r=12}(x),\text{Conv}_{r=18}(x),\text{GAP}(x)\right]; Conv_r=k(x) is the dilated convolution output with a dilation rate of k, and GAP(x) is global average pooling, compressing a 1x1 spatial dimension. Decoder: Fuses shallow (high resolution) and deep (strong semantic) features \text{Output}=\text{Conv}_{1\times1}\left(\text{Concat}\left[\text{UpSample}(\text{ASPP}(x)),\text{Conv}_{3\times3}(\text{LowLevelFeat})\right]\right) UpSample is an upsampling operation (such as bilinear interpolation) that restores resolution. LowLevelFeat is the shallow features of the backbone network (high resolution, weak semantics). Conv_{1\times1} is a 1x1 convolution that adjusts the number of channels and combines deep semantic information (ASPP output) with shallow details (low-level features) to improve segmentation accuracy.

[0031] Output processing and defect localization: (1) Model Output Output size: (H times W times C) ((C) is the number of defect categories).

[0032] The class probability at each pixel position (i,j): P_{i,j,c}=\text{Softmax}(z_{i,j,c})=\frac{e^{z_{i,j,c}}}{\sum_{c'}e^{z_{i,j,c'}}} z_{i,j,c}\): The model's original output (logits) for category \(c); P_{i,j,c} is the probability (after normalization) that pixel (i,j) belongs to category c, and c' is all possible categories (such as background, crack, pit, etc.).

[0033] (2) Defect location

[0034] Pixel-by-pixel classification: The class with the highest probability is used as the prediction result: \text{Class}_{i,j}=\arg\max_{c}P_{i,j,c}

[0035] For each pixel (i,j), select the category c with the highest probability as the prediction result.

[0036] Binary mask: Generate a binary mask for a specific defect category (c) (such as a crack): \text{Mask}_{c}[i,j]=\begin{cases} 1&\text{if}\text{Class}_{i,j}=c\text{and}P_{i,j,c}>\tau\\ 0&\text{otherwise} \end{cases} (\tau): A probability threshold (e.g., 0.5) used to filter low-confidence predictions. \text{Mask}_c: Only retain pixels of category c with a probability higher than τ.

[0037] In one embodiment, suppose we want to detect appearance defects in a ceramic vase. Possible defect categories include cracks and dents, with the background as one category, i.e. (C=3) (background, cracks, dents).

[0038] 1. Image Input

[0039] Input the image of the ceramic vase into the model. The image size is H×W.

[0040] 2. Backbone Network Feature Extraction

[0041] The ResNet-101 backbone network extracts features from the input image, obtaining feature maps at different levels and scales. During feature extraction, dilated convolutions are used to expand the receptive field and avoid resolution loss.

[0042] For example, in a feature map of a certain layer, by using dilated convolution (dilation rate r=2), the original 3x3 convolution kernel can cover a larger area, thereby capturing richer contextual information.

[0043] 3. The ASPP module captures context information.

[0044] The ASPP module processes the feature maps output by the backbone network, using dilated convolutions with dilation rates r=6, 12, and 18, and global average pooling, respectively.

[0045] Suppose that at a certain pixel location, a convolutional kernel with a dilation rate r=6 captures the local details of the crack, a convolutional kernel with a dilation rate r=18 obtains the global context information around the crack, and global average pooling provides global information for the entire vase image. Finally, all this information is concatenated.

[0046] 4. Decoder fusion features

[0047] The feature maps output by the ASPP module are upsampled to match the resolution of the shallow feature maps of the backbone network.

[0048] Assuming the shallow feature map contains some texture details of the vase surface, the upsampled ASPP output feature map is concatenated with the shallow feature map, and then the number of channels is adjusted by a 1×1 convolution to obtain the final output feature map.

[0049] Output processing and defect localization

[0050] Category probability calculation: For each pixel position (i, j) in the output feature map, calculate the probabilities that it belongs to the three categories of background, crack, and pit. For example, for a certain pixel position, the original outputs are \(z_{i_1,j_1,\text{background}} = 1.2\), \(z_{i_1,j_1,\text{crack}} = 3.5\), and \(z_{i_1,j_1,\text{pit}} = 0.8\). After calculation by the Softmax function, we get: \(P_{i_1,j_1,\text{background}}=\frac{e^{1.2}}{e^{1.2}+e^{3.5}+e^{0.8}}\approx0.09\); \(P_{i_1,j_1,\text{crack}}=\frac{e^{3.5}}{e^{1.2}+e^{3.5}+e^{0.8}}\approx0.87\); \(P_{i_1,j_1,\text{pit}}=\frac{e^{0.8}}{e^{1.2}+e^{3.5}+e^{0.8}}\approx0.04\); Pixel-by-pixel classification. Since \(P_{i_1,j_1,\text{crack}}\) is the largest, the predicted category for this pixel position is crack, that is, \(\text{Class}_{i_1,j_1}=\text{crack}\).

[0051] On the other hand, a machine vision-based automatic ceramic appearance detection system is provided, which is applied to the above-mentioned machine vision-based ceramic appearance automatic detection method. The system includes: Illumination module: A combination of multiple types of light sources is used to meet the lighting requirements at different angles and in different ways; Light source controller: Used to adjust the parameters of the light source, and at the same time, according to the different ceramic materials, dynamically adjust the light source intensity through the light source controller to obtain the best lighting effect; Sensor and feedback device: Install a light sensor to monitor the light conditions of the lighting environment in real time and feed the data back to the light source controller; Image acquisition module: Acquire ceramic images and mark the time stamp and light source type for each frame of image; Fusion module: Match its multi-light source images through feature points and perform weighted fusion to obtain a fused image; Output module: Input the fused image into the semantic segmentation model and output the defect category and location.

[0052] In this embodiment, a multi-source light source combination of a ring light source, a coaxial light source, and a low-angle linear light source is employed, and the light source parameters are optimized to illuminate the ceramic from different angles and directions. The ring light source provides uniform ring illumination, which helps to highlight the overall features of the ceramic surface; the coaxial light source reduces interference from reflected light and clearly displays minor defects on the ceramic surface; and the low-angle linear light source highlights surface irregularities. Through time-division and synchronous triggering timing control, images under different lighting conditions can be acquired, enriching image information and improving the display effect of defects. Simultaneously, the light source intensity is dynamically adjusted according to the ceramic material, and real-time adjustment is achieved using PID feedback control, enabling the lighting effect to adapt to different ceramic materials. Suitable lighting conditions can be provided for ceramic products with different surface gloss levels and color depths, ensuring that defects are clearly displayed, thus improving the adaptability and accuracy of the detection.

[0053] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0054] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for automatic detection of ceramic appearance based on machine vision, characterized in that, Includes the following steps: S1: Set up a combination of multiple light sources for illumination, optimize the light source parameters, and dynamically adjust the light source intensity according to the ceramic material; S2: Acquire ceramic images and simultaneously mark each frame with a timestamp and light source type; S3: The multi-source images are matched by feature points and then weighted and fused to obtain a fused image; S4: Input the fused image into the semantic segmentation model, and output the defect category and location.

2. The automatic detection method of ceramic appearance based on machine vision according to claim 1, characterized in that, The multi-light source combination includes: a ring light source, a coaxial light source, and a low-angle linear light source.

3. The automatic ceramic appearance detection method based on machine vision according to claim 1, characterized in that, The optimization of the light source parameters also includes: Optimize the angles of multiple light sources separately: ring light source angle: 30-45°, low-angle light source angle: 10-20°; Multi-source timing control: Time-sharing triggering: Switching light sources at different times to capture images under different lighting conditions; Synchronous triggering: Multiple light sources work simultaneously, acquiring images from multiple light sources at the same time.

4. The automatic detection method of ceramic appearance based on machine vision according to claim 1, characterized in that, The method of dynamically adjusting the light source intensity based on the ceramic material includes: using PID feedback control to adjust the light source intensity in real time. ； in It is the error between the current image grayscale variance and the target variance. , , These are PID coefficients. It is a differential term.

5. The automatic ceramic appearance inspection method based on machine vision according to claim 1, characterized in that, Feature point matching is used to include the following in its multi-light source image: Align multi-source images to ensure accurate pixel correspondence during subsequent fusion, detect key points, and generate 128-dimensional feature descriptors for feature matching: ； in It is a feature descriptor of the reference image. It is the nearest neighbor feature descriptor of the image to be matched. It is the second nearest neighbor feature descriptor of the image to be matched. It is a threshold.

6. The automatic detection method of ceramic appearance based on machine vision according to claim 1, characterized in that, The weighted fusion results in a fused image that includes: adaptive fusion of multi-source images, highlighting defect features and suppressing noise. ； wherein, is the image under the kth light source, is the weight of the kth light source, n is the number of light sources; Weight calculation based on signal-to-noise ratio: ， ； wherein is the gray scale standard deviation of the defect area, is the gray scale standard deviation of the background area.

7. The automatic detection method of ceramic appearance based on machine vision according to claim 1, characterized in that, The semantic segmentation model includes: a backbone network, an ASPP module, and a decoder; Multi-scale features are extracted from the backbone network while dilated convolution is performed to avoid resolution loss due to downsampling. The ASPP module uses parallel multi-branch dilated convolution to capture multi-scale contextual information. The decoder fuses shallow and deep features.

8. A machine vision-based automatic detection system for ceramic appearance, characterized by, The system applied to the automatic ceramic appearance inspection method based on machine vision according to any one of claims 1-7, the system comprising: Lighting module: Employs a combination of various light sources to meet lighting needs from different angles and in different ways; Light source controller: Used to adjust the parameters of the light source, and dynamically adjusts the light source intensity according to the different ceramic materials to obtain the best lighting effect; Sensor and feedback device: Equipped with light sensors to monitor the light conditions of the lighting environment in real time and feed the data back to the light source controller; Image acquisition module: Acquires ceramic images and marks each frame with a timestamp and light source type; Fusion module: It matches multi-source images using feature points and performs weighted fusion to obtain a fused image; Output module: Input the fused image into the semantic segmentation model, and output the defect category and location.

Images