A high-precision defect detection method for ceramic tiles based on multi-scale feature refinement

Abstract

The application belongs to the technical field of ceramic tile defect detection, and discloses a high-precision ceramic tile defect detection method based on multi-scale feature refinement, which comprises the following steps: collecting original images of ceramic tiles and constructing a ceramic tile defect dataset, and sequentially performing preprocessing and dataset division on the ceramic tile defect dataset; taking YOLO11n as a benchmark detection model, replacing the original down-sampling layer of YOLO11n with a CADown lightweight down-sampling module, replacing the original nearest neighbor interpolation up-sampling of YOLO11n with a DySample dynamic up-sampling operator, and replacing the original PANet feature fusion network with a BiFPN bidirectional feature pyramid network integrated with a GLSA local self-attention mechanism to obtain a CBD-YOLO detection model; and inputting the images of ceramic tiles to be detected into the trained ceramic tile defect detection model to output detection results. The application can significantly reduce the missed detection rate and the false detection rate, and has extremely lightweight and high precision.

Description

Technical Field

[0001] This invention belongs to the field of ceramic tile defect detection technology, and specifically discloses a high-precision ceramic tile defect detection method based on multi-scale feature refinement. Background Technology

[0002] The consistency of the texture on the surface of ceramic tiles is not only a key indicator of the quality of modern construction projects, but also directly determines the consistency of the product's appearance quality. However, due to the inherent random disturbances in the manufacturing process, the finished product surface inevitably produces minute defects of various shapes. Traditional quality control methods, whether relying on manual subjective visual inspection or shallow image processing technology based on edge detection and threshold segmentation, are difficult to maintain robustness in complex industrial lighting environments and high-throughput production lines, and cannot meet the stringent requirements of modern intelligent manufacturing for zero-defect delivery.

[0003] With the development of deep learning technology, the industrial quality inspection field has evolved from traditional manual visual inspection to deep learning. Two-stage detection algorithms have shown significant advantages in localization accuracy: Wang et al., based on an improved Cascade R-CNN architecture, achieved high-precision localization in complex texture backgrounds using a cascaded regression mechanism; Hu et al., by introducing a feature pyramid and attention mechanism, enhanced the ability of Mask R-CNN to perceive minute cracks. Although such algorithms have a low false detection rate, their redundant region proposal generation steps lead to significant inference latency, and the high computational cost limits their deployment in real-time, high-throughput industrial scenarios.

[0004] In contrast, single-stage algorithms improve efficiency through a single forward inference: Li et al. effectively solved the problem of high reflectivity interference on magnetic tile surfaces using an improved YOLOv5; general models such as RetinaNet and EfficientDet provide a solid theoretical foundation through Focal Loss and multi-scale fusion. However, classic single-stage detectors often suffer from high-frequency information loss during downsampling, leading to decreased detection accuracy when dealing with extremely small and vaguely defined defects due to insufficient spatial resolution of deep feature maps.

[0005] To overcome the aforementioned bottleneck in balancing speed and accuracy, recent research has primarily focused on two dimensions: lightweight architecture design and feature representation enhancement. On one hand, regarding computational efficiency optimization, Yu et al. innovatively proposed the Hybrid Vision Transformer architecture (MCAW-YOLO), utilizing a self-attention mechanism to maintain the global receptive field while reducing the number of parameters. Song et al., by improving S2C-YOLOv5, achieved efficient multi-scale defect screening under limited computing power. On the other hand, for enhancing the perception of small targets, dynamic convolution and attention mechanisms have become mainstream solutions. Wang et al.'s CTDD-YOLO introduces full-dimensional dynamic convolution (ODConv), enabling the convolution kernel to dynamically adapt to changes in input features, significantly improving the sensitivity to small targets when combined with the MPNWD loss function. An et al. attempted to integrate advanced attention modules into YOLO to address feature confusion and subtle texture localization issues in complex backgrounds.

[0006] While existing methods have made some progress, they still struggle to balance accuracy and efficiency when dealing with extremely small and low-contrast defects. Lightweight strategies often lose high-frequency details due to dimensionality reduction, weakening defect characterization; while complex attention mechanisms, although enhancing detection accuracy, significantly increase computational burden. To address these issues, it is essential to research and design a novel high-precision defect detection method for ceramic tiles based on multi-scale feature refinement, overcoming the problems inherent in existing high-precision defect detection methods for ceramic tiles. Summary of the Invention

[0007] This invention proposes a high-precision tile defect detection method based on multi-scale feature refinement to address the problems of existing high-precision tile defect detection methods, such as difficulty in balancing accuracy and efficiency, loss of high-frequency details, and weakening of defect characterization.

[0008] This invention provides a high-precision defect detection method for ceramic tiles based on multi-scale feature refinement, comprising the following steps:

[0009] S1. Collect original images of tiles and construct a tile defect dataset. Perform dual-stream multi-scale adaptive slicing preprocessing, category-aware image enhancement, and dataset partitioning on the tile defect dataset in sequence to obtain a training set, a validation set, and a test set.

[0010] S2. Using YOLO11n as the baseline detection model, the baseline model is modified by multi-scale feature refinement. The CADown lightweight downsampling module is used to replace the original downsampling layer of YOLO11n, the DySample dynamic upsampling operator is used to replace the original nearest neighbor interpolation upsampling of YOLO11n, and the BiFPN bidirectional feature pyramid network with integrated GLSA local self-attention mechanism is used to replace the original PANet feature fusion network, thus obtaining the CBD-YOLO detection model.

[0011] S3. The CBD-YOLO detection model is trained, validated, and tested using the divided training set, validation set, and test set, respectively, to obtain the trained tile defect detection model;

[0012] S4. Input the image of the tile to be detected into the tile defect detection model, and output the tile surface defect category, defect location, and confidence level detection results.

[0013] According to some embodiments of this application, a high-precision tile defect detection method based on multi-scale feature refinement is provided. In step S1, the dual-stream multi-scale adaptive slicing preprocessing includes large-scale defect ROI adaptive slicing and small-scale defect centered anchor point slicing.

[0014] The large-scale defect ROI adaptive slicing includes generating a square ROI region with a context expansion factor centered on the defect centroid, normalizing and resampling the ROI region and simultaneously transforming the defect annotation coordinates to obtain a slice set;

[0015] The small-scale defect-centered anchor point slicing includes generating candidate slices with the centroid of the unprocessed defect as the anchor point, performing boundary translation correction on image edge defects, constructing a greedy coverage screening mechanism based on the intersection-union ratio (IoA), merging overlapping slices and removing redundant samples to obtain a slice set.

[0016] According to some embodiments of this application, a high-precision tile defect detection method based on multi-scale feature refinement is provided, wherein the side length of the ROI region is... As shown in formula (1):

[0017] (1)

[0018] in, Indicates the first The width of the defect, Indicates the first The height of the defect This represents the context expansion factor.

[0019] According to some embodiments of this application, a high-precision tile defect detection method based on multi-scale feature refinement is provided. In step S1, the category-aware image enhancement includes brightness channel adaptive contrast enhancement and tail defect category geometric oversampling enhancement.

[0020] The brightness channel adaptive contrast enhancement includes converting the original image to the CIELAB color space, applying CLAHE-limited contrast adaptive histogram equalization to the brightness L channel, and enhancing the texture of low-contrast light-colored spots and defects.

[0021] The tail defect category geometric oversampling enhancement includes a geometric transformation based on the inverse category frequency to flip and rotate the sparse edge anomalies and aperture flaws tail categories in the original image, thereby balancing the category distribution.

[0022] According to some embodiments of this application, a high-precision defect detection method for ceramic tiles based on multi-scale feature refinement, in step S2, the processing flow of the CADown lightweight downsampling module includes:

[0023] After performing average pooling on the input feature map, it is divided into two parallel branches for processing. The branches include a global perception branch and a basic texture branch.

[0024] The global perception branch performs depthwise convolution and pointwise convolution operations sequentially to complete feature extraction, resulting in the output features of the global perception branch; the basic texture branch performs max pooling and basic convolution operations sequentially to complete feature extraction, resulting in the output features of the basic texture branch.

[0025] The output features of the global perception branch and the output features of the basic texture branch are channel-separated and fused to obtain a spliced ​​feature map;

[0026] A channel attention mechanism is introduced into the spliced ​​feature map. The spliced ​​feature map is then mapped to channel weights using a multilayer perceptron. Adaptive weights for each channel are generated using a sigmoid activation function. The adaptive weights are multiplied by the spliced ​​feature map to complete feature recalibration and output the final feature map, as shown in formula (2).

[0027] (2)

[0028] in, This represents the final feature map. This represents the splicing of feature maps. This represents the Sigmoid activation function. This represents a multilayer perceptron.

[0029] According to some embodiments of this application, a high-precision defect detection method for ceramic tiles based on multi-scale feature refinement, in step S2, the DySample dynamic upsampling operator processing flow includes:

[0030] The feature map and the preset upsampling ratio are obtained. The channel dimension of the input feature is reduced by a linear projection layer to generate the coordinate offset feature of the corresponding local sampling region, as shown in formula (3):

[0031] (3)

[0032] in, Indicates coordinate offset characteristics. Represents the feature map;

[0033] The coordinate offset features are reshaped by a pixel shuffling operation to obtain an offset feature map that matches the high-resolution size of the target.

[0034] A set of standard grid sampling points is constructed, and the standard grid coordinates are superimposed with the coordinate offset feature vector to generate an adaptive dynamic sampling point set, as shown in formula (4):

[0035] (4)

[0036] in, Represents the set of sampling points. Represents standard grid coordinates;

[0037] The feature map is resampled and interpolated using the grid_sample function based on the set of sampling points, and a high-resolution feature map with restored details is output, as shown in formula (5):

[0038] (5)

[0039] in, This represents a high-resolution feature map.

[0040] According to some embodiments of this application, a high-precision tile defect detection method based on multi-scale feature refinement is provided. In step S2, the BiFPN bidirectional feature pyramid network integrating GLSA local self-attention mechanism includes a pre-feature calibration unit and a weighted multi-scale fusion unit.

[0041] The pre-feature calibration unit adopts a dual-branch parallel structure of global branch and local branch. The global branch is used to capture cross-scale long-range dependence to cope with large-scale defects, and the local branch is used to enhance the high-frequency texture of small defects in the tile. The dual-branch residual fusion outputs a calibration single-scale feature map.

[0042] The weighted multi-scale fusion unit configures learnable weights for input features at each scale through cross-scale bidirectional skip connections and adopts a fast normalization fusion strategy to adaptively allocate the contribution of features at different resolutions to defect detection.

[0043] According to some embodiments of this application, a high-precision tile defect detection method based on multi-scale feature refinement, in step S2, the BiFPN bidirectional feature pyramid network processing flow integrating the GLSA local self-attention mechanism includes:

[0044] Receive the final feature map output by the backbone network through the CADown lightweight downsampling module;

[0045] Each layer of the original final feature map is fed into an independent pre-feature calibration unit to complete feature calibration. The processing flow for a single final feature map is as follows: the final feature map is split into a global branch and a local branch for parallel processing; the global branch models long-range dependencies across pixels through global self-attention operations, extracts global semantic features of large-size defects, and obtains the global branch output features; the local branch focuses on local pixel regions of the feature map through local self-attention operations, extracts high-frequency texture details of small defects, and obtains the local branch output features; the global branch output features and the local branch output features are fused using residual addition to obtain the calibrated single-scale feature map;

[0046] The calibration single-scale feature maps, after being calibrated by the pre-feature calibration unit, are input into the weighted multi-scale fusion unit to build a multi-feature fusion node with bidirectional cross-scale jump connections from top to bottom and from bottom to top. Independent learnable weights are configured for each multi-feature fusion node, and the value range of all learnable weights is constrained by the ReLU activation function. A fast normalization fusion strategy is used to weight and sum the multiple input features within the same node to complete the adaptive aggregation of defect features at different resolutions, and output a multi-scale feature map with multi-level fusion, as shown in formula (6):

[0047] (6)

[0048] in, Represents multi-scale feature maps, Indicates the sequence number of the calibration single-scale feature map. Indicates the first Learnable weights corresponding to each calibration single-scale feature map Used to maintain numerical stability Indicates the sequence number of the calibration single-scale feature map. Indicates the first Learnable weights corresponding to each calibration single-scale feature map This represents the calibration of a single-scale feature map.

[0049] According to some embodiments of this application, a high-precision tile defect detection method based on multi-scale feature refinement is provided, wherein in step S1, the data ratio of the training set, validation set, and test set is 8:1:1.

[0050] According to some embodiments of this application, a high-precision tile defect detection method based on multi-scale feature refinement is provided, wherein the defect categories include edge anomalies, corner anomalies, white spot defects, halo defects, black spot defects, and shallow spot defects.

[0051] This invention proposes a high-precision defect detection method for ceramic tiles based on multi-scale feature refinement. It introduces a lightweight CADown downsampling module, reconstructing it based on ADown, using depthwise separable convolution to replace redundant computation, and embedding a channel attention mechanism to enhance feature selection. This design significantly reduces the number of parameters and FLOPs while maximizing the preservation of high-frequency texture information on the tile surface through an optimized feature aggregation strategy. A BiFPN bidirectional feature pyramid network integrating a GLSA local self-attention mechanism is designed to replace the original PANet. It utilizes cross-scale connections and a local-global self-attention mechanism to enhance the extraction and fusion of multi-scale, especially fuzzy, defect features. The DySample dynamic upsampling operator is employed, requiring no additional parameter learning. It dynamically generates upsampling kernels based on the input, avoiding the information smoothing effect of traditional interpolation and accurately restoring the geometric details of minute defects. This method can meet industrial-grade real-time requirements while preserving micro-texture and reducing computational load. It also effectively overcomes interference from complex background textures, significantly reducing the false negative and false positive rates. Attached Figure Description

[0052] Figure 1 This is a flowchart illustrating a high-precision defect detection method for ceramic tiles based on multi-scale feature refinement according to the present invention.

[0053] Figure 2 This is a schematic diagram of the overall architecture of the CBD-YOLO detection model of the present invention;

[0054] Figure 3 The diagrams show the structures of the ADown module and the CADown lightweight downsampling module. (a) is a schematic diagram of the ADown module, and (b) is a schematic diagram of the CADown lightweight downsampling module.

[0055] Figure 4 Here are schematic diagrams of the structures of PANet and BiFPN. (a) is a schematic diagram of the structure of PANet, and (b) is a schematic diagram of the structure of the lightweight downsampling module of BiFPN.

[0056] Figure 5 This is a schematic diagram of the tile defect types in Embodiment 3 of the present invention;

[0057] Figure 6 This is a distribution diagram of the size range of defects on the surface of the ceramic tile in Embodiment 3 of the present invention;

[0058] Figure 7 This is a schematic diagram of the precision-efficiency trade-off analysis in the ablation experiment of Embodiment 3 of the present invention;

[0059] Figure 8 This is a qualitative comparative analysis diagram of the detection of minute defects in complex scenarios in Embodiment 3 of the present invention;

[0060] Figure 9 This is a radar chart illustrating the enhanced confidence level for multi-category defect detection in Embodiment 3 of the present invention. Detailed Implementation

[0061] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and should not be construed as limiting the scope of the invention.

[0062] Example 1: This example provides a high-precision defect detection method for ceramic tiles based on multi-scale feature refinement, such as... Figure 1 As shown, it includes the following steps:

[0063] S1. Collect original images of tiles and construct a tile defect dataset. Perform dual-stream multi-scale adaptive slicing preprocessing, category-aware image enhancement, and dataset partitioning on the tile defect dataset in sequence to obtain training set, validation set, and test set.

[0064] S2. Using YOLO11n as the baseline detection model, the baseline model is modified by multi-scale feature refinement. The CADown lightweight downsampling module is used to replace the native downsampling layer of YOLO11n, the DySample dynamic upsampling operator is used to replace the native nearest neighbor interpolation upsampling of YOLO11n, and the BiFPN bidirectional feature pyramid network with integrated GLSA local self-attention mechanism is used to replace the native PANet feature fusion network, thus obtaining the CBD-YOLO detection model.

[0065] S3. The CBD-YOLO detection model is trained, validated, and tested using the divided training, validation, and test sets respectively, to obtain the trained tile defect detection model.

[0066] S4. Input the image of the tile to be detected into the tile defect detection model, and output the tile surface defect category, defect location, and confidence level detection results.

[0067] Example 2: This example provides a high-precision defect detection method for ceramic tiles based on multi-scale feature refinement, including the following steps:

[0068] S1. Collect original images of tiles and construct a tile defect dataset. Perform dual-stream multi-scale adaptive slicing preprocessing, category-aware image enhancement, and dataset partitioning on the tile defect dataset in sequence to obtain training set, validation set, and test set.

[0069] Specifically, facing the challenge of coexisting defects with extreme scale differences in ultra-high resolution images, standard uniform sliding window slicing methods often suffer from two shortcomings: they easily disrupt the integrity of large-scale defects and lead to excessive background noise in small-scale defect samples. To address these limitations, this application proposes Dual-Stream Multi-Scale Adaptive Slicing (DS-MSAS), which dynamically splits the processing based on the spatial scale characteristics of defects. DS-MSAS includes adaptive slicing of large-scale defect ROIs and centered anchor point slicing of small-scale defects.

[0070] In this embodiment, defects exceeding 640×640 pixels are defined as large-scale defects. An adaptive Region of Interest (ROI) generation mechanism is introduced to maintain the integrity of its topological structure. The adaptive slicing of the large-scale defect ROI involves generating a square ROI region centered on the defect centroid with a context expansion factor. The ROI region is then normalized and resampled, and the defect annotation coordinates are simultaneously transformed to obtain a slice set. The ROI region is extracted and normalized to the standard input size using bilinear interpolation. This adaptive resampling ensures that the global geometry of the large-scale defect is completely preserved in a single inference view, while the coordinate mapping function correspondingly transforms the annotation information. The side length of the ROI region... As shown in formula (1):

[0071] (1)

[0072] in, Indicates the first The width of the defect, Indicates the first The height of the defect This represents the context expansion factor.

[0073] To address small-scale defects, this paper proposes a strategy combining centralized anchor point slicing and greedy coverage optimization, aiming to solve the problems of low proportion of small samples in the original image and background redundancy. Centralized anchor point slicing for small-scale defects involves generating candidate slices using the centroid of unprocessed defects as anchor points, performing boundary translation correction on image edge defects, constructing a greedy coverage filtering mechanism based on the Intersection over Area (IoA), merging overlapping slices, and removing redundant samples to obtain a slice set. Specifically, in this embodiment, during the slice generation stage, the strategy uses unprocessed defects as core anchor points. For each defect target, a candidate slice window is generated that is forced to align with its centroid. For defects located at the image edge, a boundary adaptive correction mechanism is introduced, translating the coordinates inward while maintaining a constant slice size, ensuring the slice is completely within the image domain and avoiding feature interference caused by invalid padding. Secondly, to suppress computational redundancy caused by dense defect clusters, a sample filtering mechanism based on a greedy strategy is constructed. The coverage function is defined as the Intersection over Area (IoA) of adjacent defects within the current slice. The optimization process employs an iterative greedy algorithm: once a centered slice is selected, its coverage of neighboring defects is calculated; any defect with an IoA exceeding a preset threshold is marked as "processed" and no further slices are generated. Theoretically, this method guarantees that each small defect is prominent in at least one training sample, while significantly compressing the redundant volume of the training set.

[0074] To mitigate the inherent class imbalance and low contrast characteristics of specific defect categories, this embodiment integrates category-aware image enhancement into the data preprocessing workflow. Category-aware image enhancement includes adaptive contrast enhancement for the brightness channel and geometric oversampling enhancement for tail defect categories.

[0075] To address the issue of light-colored block defects with low signal-to-noise ratio and weak features, this embodiment employs luminance channel adaptive contrast enhancement. Luminance channel adaptive contrast enhancement includes converting the original image to the CIELAB color space, applying CLAHE-limited contrast adaptive histogram equalization to the luminance L channel, and enhancing the texture of low-contrast light-colored spot defects. This selective enhancement strategy effectively amplifies local texture contrast while avoiding distortion of chromaticity information.

[0076] For tail categories with extremely scarce samples, such as aperture defects and edge anomalies, this embodiment uses tail defect category geometric oversampling enhancement. Tail defect category geometric oversampling enhancement includes a geometric transformation based on the inverse class frequency to flip and rotate the scarce edge anomalies and aperture defects tail categories in the original image, balancing the class distribution, effectively smoothing the class distribution of the dataset, and improving the model's robustness to class imbalance.

[0077] To prevent different slices of the same image from appearing in both the training and test sets simultaneously, the principle of "segmenting the original image first, then slicing" is followed. The specific process is as follows: First, a unique identifier list for all original images is created and randomly shuffled. Then, the original image is divided into training, validation, and test sets in an 8:1:1 ratio. In the subsequent slice generation stage, all slices inherit the dataset attributes of their respective original images and are automatically distributed to the corresponding directory structure. The final generated annotation file includes category indices and normalized bounding box coordinates, fully conforming to the standard input specifications of the YOLO series algorithms.

[0078] S2. In this embodiment, YOLO11 has undergone a significant upgrade in its feature extraction architecture. YOLO11 integrates the C3k2 module, SPPF, and C2PSA components, enhancing the model's ability to capture complex visual features. The YOLO11 series offers various specifications, from lightweight (YOLOv11n) to large-scale (YOLOv11x), to adapt to different computing power scenarios. This application primarily targets resource-constrained embedded devices and has strict requirements for real-time detection speed in industrial settings; therefore, this embodiment selects YOLO11n as the baseline model. Compared to other versions, YOLO11n significantly reduces the number of parameters and computational resource consumption while maintaining competitive detection accuracy, achieving a good trade-off between inference speed and accuracy, and providing an ideal infrastructure for subsequent optimization and improvement.

[0079] To address the challenges of small size, irregular shape, and low background contrast in tile surface defects, this embodiment proposes a CBD-YOLO detection model based on YOLO11. Using YOLO11n as the baseline detection model, the model undergoes multi-scale feature refinement. The CADown lightweight downsampling module replaces the native downsampling layer of YOLO11n, introducing depthwise separable convolution and channel attention to effectively preserve the edge texture of irregular defects while reducing computational cost. The DySample dynamic upsampling operator replaces the native nearest-neighbor interpolation upsampling of YOLO11n, suppressing feature aliasing and enhancing the reconstruction clarity of small defects. A BiFPN bidirectional feature pyramid network integrating GLSA local self-attention mechanism replaces the native PANet feature fusion network, strengthening multi-scale feature weighted fusion and improving the ability to detect low-contrast defects. The resulting CBD-YOLO detection model has the following overall architecture: Figure 2 As shown.

[0080] Detection of surface defects in ceramic tiles not only faces the challenges of complex and varied defect morphologies and high localization, but also places extremely high demands on real-time detection in production environments. Traditional downsampling operations, such as stride convolution, often result in the loss of minute or low-contrast defect details while reducing feature map resolution to expand the receptive field. Therefore, this embodiment proposes an improved lightweight downsampling module using CADown, while retaining the advantages of the dual-branch pooling of the ADown module. This module aims to reduce the number of model parameters while enhancing the ability to preserve subtle defect features. The CADown lightweight downsampling module is a reconstruction based on the ADown module, such as... Figure 3 As shown, the CADown lightweight downsampling module continues the parallel architecture of channel segmentation after average pooling of input features. The CADown lightweight downsampling module introduces depthwise separable convolution in the global perception branch, decoupling standard convolution into independent depthwise convolution and pointwise convolution. This improvement significantly reduces computational complexity and the number of parameters while maintaining the original receptive field. Secondly, the CADown lightweight downsampling module embeds a channel attention mechanism after feature fusion in the two branches. The channel attention mechanism establishes a dependency model between channels, assigning adaptive weights to each feature channel, enabling the network to autonomously focus on key channels containing defect information and suppress background noise interference. Through these improvements, the CADown lightweight downsampling module can dynamically enhance feature saliency while maintaining a lightweight model, thereby effectively improving the detection accuracy of small tile defects.

[0081] Specifically, the CADown lightweight downsampling module processing flow includes:

[0082] After performing average pooling on the input feature map, it is divided into two parallel branches for processing. The branches include a global perception branch and a basic texture branch.

[0083] The global perception branch performs depthwise convolution and pointwise convolution operations sequentially to complete feature extraction, resulting in the output features of the global perception branch; the basic texture branch performs max pooling and basic convolution operations sequentially to complete feature extraction, resulting in the output features of the basic texture branch.

[0084] The output features of the global perception branch and the output features of the basic texture branch are concatenated and fused to obtain a concatenated feature map.

[0085] A channel attention mechanism is introduced into the concatenated feature map. The channel weights of the concatenated feature map are mapped by a multilayer perceptron, and then adaptive weights for each channel are generated by the sigmoid activation function. The adaptive weights are multiplied by the concatenated feature map to complete the feature recalibration and output the final feature map, as shown in formula (2):

[0086] (2)

[0087] in, This represents the final feature map. This represents the splicing of feature maps. This represents the Sigmoid activation function. This represents a multilayer perceptron.

[0088] In ceramic tile surface defect detection, defects often exhibit characteristics of being small in scale and highly irregular in shape. YOLO11's default nearest neighbor interpolation can easily lead to the loss of high-frequency details of small defects and limited positioning accuracy. This embodiment introduces the DySample dynamic upsampling operator. The DySample dynamic upsampling operator is based on content-aware point sampling and uses a lightweight linear projection (LP) mode to replace the expensive dynamic convolution, achieving efficient feature deformation correction while meeting industrial real-time requirements.

[0089] The DySample dynamic upsampling operator processing flow includes:

[0090] The feature map and the preset upsampling ratio are obtained. The channel dimension of the input feature is reduced by a linear projection layer to generate the coordinate offset feature of the corresponding local sampling region, as shown in formula (3):

[0091] (3)

[0092] in, Indicates coordinate offset characteristics. Represents the feature map.

[0093] A pixel shuffling operation is used to reshape the coordinate offset features, resulting in an offset feature map that matches the high-resolution size of the target.

[0094] A set of standard grid sampling points is constructed, and the standard grid coordinates are superimposed with the coordinate offset feature vector to generate an adaptive dynamic sampling point set, as shown in formula (4):

[0095] (4)

[0096] in, Represents the set of sampling points. Represents standard grid coordinates.

[0097] The grid_sample function is used to resample and interpolate the feature map based on the set of sampling points, and outputs a high-resolution feature map with restored details, as shown in formula (5):

[0098] (5)

[0099] in, This represents a high-resolution feature map.

[0100] Through the aforementioned mechanism, the DySample dynamic upsampling operator can learn the spatial offset of each sampling point in a targeted manner, enabling dynamic adjustment of feature map details. This point-level adaptive capability not only more accurately recovers the details of surface defects in tiles but also significantly enhances the model's ability to perceive and locate small targets.

[0101] To address the issues of feature weight uniformity and missing context dependencies in the original YOLO11 PANet architecture when handling multi-scale defects, this embodiment proposes a BiFPN bidirectional feature pyramid network integrating a GLSA local self-attention mechanism to replace the native PANet feature fusion network. This embodiment uses BiFPN to replace PANet to achieve adaptive weighted aggregation of features, such as... Figure 6 As shown, compared to PANet Figure 4 (a) Simple one-way path, BiFPN Figure 4 (b) The network topology was optimized by adding cross-scale skip connections, breaking the limitation of feature equalization. Its core fast normalization fusion strategy assigns a learnable weight to each input, enabling the network to dynamically adjust the contribution of different resolutions according to task requirements.

[0102] The BiFPN bidirectional feature pyramid network integrating the GLSA local self-attention mechanism includes a pre-feature calibration unit and a weighted multi-scale fusion unit.

[0103] The pre-feature calibration unit adopts a dual-branch parallel structure with global and local branches. The global branch is used to capture long-range dependencies across scales to cope with large-scale defects, while the local branch is used to enhance the high-frequency texture of small defects in the tile. The dual-branch residual fusion outputs a calibration single-scale feature map.

[0104] The weighted multi-scale fusion unit assigns learnable weights to input features at each scale through bidirectional skip connections across scales and adopts a fast normalization fusion strategy to adaptively allocate the contribution of features at different resolutions to defect detection.

[0105] The processing flow of the BiFPN bidirectional feature pyramid network integrating the GLSA local self-attention mechanism includes:

[0106] Receive the final feature map output by the backbone network via the CADown lightweight downsampling module.

[0107] Each layer of the original final feature map is fed into an independent pre-feature calibration unit to complete feature calibration. The processing flow of a single final feature map is as follows: the final feature map is split into two parallel processing paths: a global branch and a local branch. The global branch models long-range dependencies across pixels through global self-attention operations, extracts global semantic features of large-size defects, and obtains the global branch output features. The local branch focuses on local pixel regions of the feature map through local self-attention operations, extracts high-frequency texture details of small defects, and obtains the local branch output features. The global branch output features and the local branch output features are fused using residual addition to obtain the calibrated single-scale feature map.

[0108] The calibration single-scale feature maps calibrated by the pre-feature calibration unit are input into the weighted multi-scale fusion unit to build a multi-feature fusion node with bidirectional cross-scale jump connections from top to bottom and from bottom to top; each multi-feature fusion node is configured with an independent learnable weight, and all learnable weights are constrained in their range by the ReLU activation function; a fast normalization fusion strategy is used to weight and sum the multiple input features in the same node to complete the adaptive aggregation of defect features at different resolutions, and output the multi-scale feature map with multi-level fusion, as shown in formula (6):

[0109] (6)

[0110] in, Represents multi-scale feature maps, Indicates the sequence number of the calibration single-scale feature map. Indicates the first Learnable weights corresponding to each calibration single-scale feature map Used to maintain numerical stability Indicates the sequence number of the calibration single-scale feature map. Indicates the first Learnable weights corresponding to each calibration single-scale feature map This represents the calibration of a single-scale feature map.

[0111] The BiFPN bidirectional feature pyramid network, which integrates the GLSA local self-attention mechanism, employs a synergistic mechanism of first calibration (GLSA) and then weighting (BiFPN), which can preserve high-resolution semantic information that is beneficial for locating small irregular defects to the greatest extent without significantly increasing computational costs.

[0112] S3. The CBD-YOLO detection model is trained, validated, and tested using the divided training, validation, and test sets respectively, to obtain the trained tile defect detection model.

[0113] S4. Input the image of the tile to be detected into the tile defect detection model, and output the tile surface defect category, defect location, and confidence level detection results.

[0114] Specifically, the defect categories include edge abnormalities, corner abnormalities, white spot blemishes, aperture blemishes, black spot blemishes, and light spot blemishes.

[0115] Example 3: In this example, the data was collected from a well-known tile manufacturing company in Foshan City, Guangdong Province, and includes 5388 original tile images. The image resolution specifications are mainly [specifications to be filled in]. Pixels and The dataset features extremely high pixel density, typical of ultra-high resolution industrial inspection scenarios. It covers six typical defect categories: Edge Anomaly, Corner Anomaly, White Spot Blemish, Aperture Blemish, Dark Dot Blemish, and Light Spot Blemish. Typical samples include... Figure 5 As shown.

[0116] Figure 6 This paper presents the statistical results of the scale distribution of 15,230 defect samples in this ceramic tile defect dataset. Analysis shows that the data distribution exhibits significant scale imbalance. This extreme proportion of small targets poses a serious challenge to existing general-purpose detection architectures: mainstream convolutional neural networks, represented by YOLO11, are prone to feature dispersion or even complete loss of subtle defect textures during deep feature extraction due to their continuous downsampling operations. This irreversible attenuation of semantic information directly limits the robustness and precision detection capabilities of algorithms in real-world industrial scenarios.

[0117] This embodiment adopts the method of Embodiment 2, and all experiments in this embodiment are conducted on a computing platform based on the Windows 11 operating system. The software environment is based on the Python 3.10.16 programming language, and utilizes the PyTorch 2.6.0 deep learning framework combined with the CUDA 12.6 computing library for model building and training. For hardware acceleration, a workstation equipped with an NVIDIA GeForce RTX 4070 SUPER graphics processing unit (GPU) is used, which is equipped with 12282 MiB of video memory to meet the computing power requirements of high-resolution image processing. Specific hyperparameter configurations during the model training phase are detailed in Table 1.

[0118] Table 1 Experimental Parameter Settings

[0119]

[0120] To comprehensively quantify and improve the detection performance and effectiveness of the model, this embodiment selects precision (P), recall (R), and mean average precision (mAP) as the core indicators for measuring the ability to detect defects on tile surfaces. Simultaneously, parameters and floating-point operations (GFLOPs) are introduced to evaluate the model's computational complexity and lightweight nature.

[0121] Accuracy As shown in formula (7):

[0122] (7)

[0123] Recall rate As shown in formula (8):

[0124] (8)

[0125] Mean accuracy As shown in formula (9):

[0126] (9)

[0127] in, This represents the number of positive samples that were correctly identified as defects. This represents the number of negative samples that were mistakenly identified as defects. This represents the number of actual defects that the model missed detecting. Indicates the total number.

[0128] To verify the effectiveness of the various improvement strategies proposed in this embodiment (CADown lightweight downsampling module, DySample dynamic upsampling operator, and BiFPN bidirectional feature pyramid network integrating GLSA local self-attention mechanism), orthogonal ablation experiments were conducted using YOLO11n as the baseline model. Quantitative analysis of the experimental results is detailed in Table 2.

[0129] Table 2 Ablation Experiment Results

[0130]

[0131] First, addressing the spatial redundancy issue in high-resolution feature map downsampling, the CADown lightweight downsampling module demonstrates significant feature compression capabilities. The CADown lightweight downsampling module reduces the number of model parameters by 20%. At the same time, the average precision (mAP@0.5) was improved by 1.7 percentage points. This phenomenon indicates that traditional stride convolution introduces a large amount of invalid computation when processing tile backgrounds with simple textures, while the CADown lightweight downsampling module effectively suppresses the propagation of low-frequency background noise through adaptive channel weighting and spatial pooling, thereby enhancing the encoding efficiency of salient defect features.

[0132] Secondly, the DySample dynamic upsampling operator has proven to be key to solving the problem of aligning small defect features. Although this operator introduces a small amount of computational overhead (0.1 GFLOPs), it effectively avoids feature fuzziness caused by traditional interpolation algorithms through a content-aware point sampling mechanism, thereby improving the model precision to 79.2%. This confirms that in pixel-level defect detection, content-based dynamic reconstruction is superior to static geometric transformation.

[0133] Furthermore, the BiFPN bidirectional feature pyramid network integrating the GLSA local self-attention mechanism reshapes the fusion path of multi-scale features. By introducing a weighted bidirectional feature pyramid and a global-local self-attention mechanism, this combination significantly improves the recall rate to 75.8% while maintaining a lightweight parameter count (2.08M). This indicates that the core of solving the missed detection problem in complex industrial scenarios lies in building a global context-aware capability that can overcome scale barriers, enabling the model to simultaneously handle the feature representation of large-scale color blocks and small-scale textures.

[0134] like Figure 7 As shown, the horizontal axis represents the number of model parameters (Params), the vertical axis represents the detection accuracy (mAP@0.5), and the bubble size corresponds to the computational cost (GFLOPs). The red bubble represents the final model proposed in this paper (Ours), located in the upper left corner of the coordinate system, indicating that optimal detection performance is achieved while significantly reducing computational resource consumption. When the above three strategies are organically integrated into the same framework, a significant synergistic enhancement effect is produced. Figure 7 As shown, the final model (Ours) not only surpasses the baseline in accuracy (reaching 80.2% mAP@0.5, an improvement of 2.3%), but also achieves a breakthrough reduction in model complexity (the number of parameters is reduced by 33.6% to 1.72M, and the computational cost is reduced by 9.4% to 5.8G). This strongly confirms that the architecture proposed in this study is not a simple stacking of modules, but a compact topology deeply optimized for edge computing environments, providing a new theoretical paradigm for high-precision real-time industrial detection under resource-constrained conditions.

[0135] To further verify the comprehensive advantages of the improved algorithm, the proposed improved model CBD-YOLO was rigorously compared with current mainstream object detection algorithms (including YOLOv5n, YOLOv8n, YOLO10n, YOLO11n, and the latest YOLO12n) under the same experimental environment. As shown in Table 3, although the YOLO series algorithms have improved detection performance through continuous architectural iterations, they often struggle to achieve lightweight models, exhibiting a significant "accuracy-efficiency" bottleneck.

[0136] Table 3. Comparison of the CBD-YOLO model in this embodiment with other different models.

[0137]

[0138] Specifically, the latest YOLO12n model achieves a highly competitive accuracy of 79.8%, but its parameter count (2.57M) is relatively high. In contrast, the tile defect detection model (CBD-YOLO) proposed in this embodiment achieves an optimal detection accuracy of 80.2% using only 1.72M parameters. Compared to the benchmark model YOLO11n and other state-of-the-art (SOTA) models, CBD-YOLO achieves superior accuracy while reducing the parameter count by approximately 33%. This result strongly demonstrates that CBD-YOLO successfully breaks the traditional positive correlation between accuracy and cost, providing an optimal solution for resource-constrained industrial quality inspection tasks that combines high performance with extreme lightweight design.

[0139] To visually evaluate the model's generalization ability in extremely complex scenarios, the performance of the improved CBD-YOLO model and the baseline model (YOLO11n) on typical difficult samples was visually compared, such as... Figure 8 As shown in the figure. To visually evaluate the model's feature extraction capability against extremely complex texture backgrounds, representative samples of small and low-contrast defects from the test set were selected for visualization analysis. The comparison results are shown in the figure. Figure 8 As shown in the figure, the significant differences between the baseline model (YOLO11n) and the CBD-YOLO proposed in this embodiment in handling fine features are illustrated. In the detection task targeting "Dark Dot Blemish" (…), Figure 8In the first and second columns, due to the extremely small size of the defect targets (<32×32 pixels) and their high integration with the tile surface texture, the benchmark model exhibits significant perceptual limitations, leading to severe missed detections or only being able to locate them with extremely low confidence (0.4-0.6). This uncertainty is unacceptable in industrial quality inspection. In contrast, thanks to the DySample dynamic upsampling operator's ability to preserve high-frequency details, CBD-YOLO successfully captured all the tiny defects and significantly improved the detection confidence to above 0.9. Furthermore, when processing "Corner Anomalies" located at image boundaries (… Figure 8 (Third column) CBD-YOLO effectively mitigates the edge truncation effect through its improved feature pyramid structure, raising the confidence level from 0.5 to 0.9, demonstrating that it still maintains extremely high localization robustness in non-central areas.

[0140] To further quantify the model's generalization performance across different defect categories, the average detection confidence for each category on the test set was statistically analyzed, and radar comparison charts were plotted, such as... Figure 9 As shown in the figure, the overall performance envelope of CBD-YOLO (red solid line area) relative to the baseline model YOLO11n (gray dashed line area) is clearly depicted. Geometrically, the confidence polygon of CBD-YOLO completely covers the baseline model, indicating that the improved algorithm outperforms the baseline model across all examined defect categories. Particularly noteworthy is the near doubling of the model's average confidence on the traditionally long-tailed difficult examples "Light Spot" and "Edge Anomaly," increasing from 0.5 to 0.9. This breakthrough in data dimension demonstrates the crucial role of the GLSA global attention mechanism in suppressing background noise and enhancing the saliency of weak features. In summary, CBD-YOLO not only solves the problem of "difficult detection" but also sets a new benchmark in terms of "accurate detection," providing solid algorithmic support for achieving zero-miss detection in automated industrial quality inspection.

[0141] This application focuses on the challenge of detecting minute and low-contrast defects on ceramic tile surfaces in industrial settings, and proposes a detection algorithm that combines extreme lightweight design with high accuracy. By introducing the CADown lightweight downsampling module to reconstruct the backbone network downsampling process, and combining the DySample dynamic upsampling operator with a BiFPN bidirectional feature pyramid network integrating GLSA local self-attention mechanism, this application solves the key bottleneck of losing high-frequency details during the dimensionality reduction process of the lightweight model.

[0142] Ablation experiments and comparative analysis confirm that the improved algorithm achieves a 2.3% improvement in detection accuracy while significantly reducing the number of model parameters by 33.6%. Visualization results further demonstrate that CBD-YOLO can effectively overcome interference from complex background textures, significantly reducing both false negative and false positive rates. This application not only provides an efficient industrial vision inspection model but also offers a theoretical paradigm of "redundancy removal and feature preservation" for deep learning model design in edge computing environments, showing broad application prospects in the field of intelligent manufacturing quality inspection.

[0143] The embodiments of the present invention are given for illustrative and descriptive purposes only, and are not intended to be exhaustive or to limit the invention to the forms disclosed. Many modifications and variations will be apparent to those skilled in the art. The embodiments were chosen and described in order to better illustrate the principles and practical application of the invention, and to enable those skilled in the art to understand the invention and to design various embodiments with various modifications suitable for a particular purpose.

Claims

1. A high-precision defect detection method for ceramic tiles based on multi-scale feature refinement, characterized in that, Includes the following steps: S1. Collect original images of tiles and construct a tile defect dataset. Perform dual-stream multi-scale adaptive slicing preprocessing, category-aware image enhancement, and dataset partitioning on the tile defect dataset in sequence to obtain a training set, a validation set, and a test set. S2. Using YOLO11n as the baseline detection model, the baseline model is modified by multi-scale feature refinement. The CADown lightweight downsampling module is used to replace the original downsampling layer of YOLO11n, the DySample dynamic upsampling operator is used to replace the original nearest neighbor interpolation upsampling of YOLO11n, and the BiFPN bidirectional feature pyramid network with integrated GLSA local self-attention mechanism is used to replace the original PANet feature fusion network, thus obtaining the CBD-YOLO detection model. S3. The CBD-YOLO detection model is trained, validated, and tested using the divided training set, validation set, and test set, respectively, to obtain the trained tile defect detection model; S4. Input the image of the tile to be detected into the tile defect detection model, and output the tile surface defect category, defect location, and confidence level detection results.

2. The high-precision defect detection method for ceramic tiles based on multi-scale feature refinement according to claim 1, characterized in that, In step S1, the dual-stream multi-scale adaptive slicing preprocessing includes large-scale defect ROI adaptive slicing and small-scale defect centered anchor point slicing. The large-scale defect ROI adaptive slicing includes generating a square ROI region with a context expansion factor centered on the defect centroid, normalizing and resampling the ROI region and simultaneously transforming the defect annotation coordinates to obtain a slice set; The small-scale defect-centered anchor point slicing includes generating candidate slices with the centroid of the unprocessed defect as the anchor point, performing boundary translation correction on image edge defects, constructing a greedy coverage screening mechanism based on the intersection-union ratio (IoA), merging overlapping slices and removing redundant samples to obtain a slice set.

3. The high-precision defect detection method for ceramic tiles based on multi-scale feature refinement according to claim 2, characterized in that, The side length of the ROI region As shown in formula (1): （1） in, Indicates the first The width of the defect, Indicates the first The height of the defect This represents the context expansion factor.

4. The high-precision defect detection method for ceramic tiles based on multi-scale feature refinement according to claim 1, characterized in that, In step S1, the category-aware image enhancement includes brightness channel adaptive contrast enhancement and tail defect category geometric oversampling enhancement; The brightness channel adaptive contrast enhancement includes converting the original image to the CIELAB color space, applying CLAHE-limited contrast adaptive histogram equalization to the brightness L channel, and enhancing the texture of low-contrast light-colored spots and defects. The tail defect category geometric oversampling enhancement includes a geometric transformation based on the inverse category frequency to flip and rotate the sparse edge anomalies and aperture flaws tail categories in the original image, thereby balancing the category distribution.

5. The high-precision defect detection method for ceramic tiles based on multi-scale feature refinement according to claim 1, characterized in that, In step S2, the CADown lightweight downsampling module processing flow includes: After performing average pooling on the input feature map, it is divided into two parallel branches for processing. The branches include a global perception branch and a basic texture branch. The global perception branch performs depthwise convolution and pointwise convolution operations sequentially to complete feature extraction, resulting in the output features of the global perception branch; the basic texture branch performs max pooling and basic convolution operations sequentially to complete feature extraction, resulting in the output features of the basic texture branch. The output features of the global perception branch and the output features of the basic texture branch are channel-separated and fused to obtain a spliced ​​feature map; A channel attention mechanism is introduced into the spliced ​​feature map. The spliced ​​feature map is then mapped to channel weights using a multilayer perceptron. Adaptive weights for each channel are generated using a sigmoid activation function. The adaptive weights are multiplied by the spliced ​​feature map to complete feature recalibration and output the final feature map, as shown in formula (2). （2） in, This represents the final feature map. This represents the splicing of feature maps. This represents the Sigmoid activation function. This represents a multilayer perceptron.

6. The high-precision defect detection method for ceramic tiles based on multi-scale feature refinement according to claim 5, characterized in that, In step S2, the DySample dynamic upsampling operator processing flow includes: The feature map and the preset upsampling ratio are obtained. The channel dimension of the input feature is reduced by a linear projection layer to generate the coordinate offset feature of the corresponding local sampling region, as shown in formula (3): （3） in, Indicates coordinate offset characteristics. Represents the feature map; The coordinate offset features are reshaped by a pixel shuffling operation to obtain an offset feature map that matches the high-resolution size of the target. A set of standard grid sampling points is constructed, and the standard grid coordinates are superimposed with the coordinate offset feature vector to generate an adaptive dynamic sampling point set, as shown in formula (4): （4） in, Represents the set of sampling points. Represents standard grid coordinates; The feature map is resampled and interpolated using the grid_sample function based on the set of sampling points, and a high-resolution feature map with restored details is output, as shown in formula (5): （5） in, This represents a high-resolution feature map.

7. The high-precision defect detection method for ceramic tiles based on multi-scale feature refinement according to claim 6, characterized in that, In step S2, the BiFPN bidirectional feature pyramid network integrating the GLSA local self-attention mechanism includes a pre-feature calibration unit and a weighted multi-scale fusion unit. The pre-feature calibration unit adopts a dual-branch parallel structure of global branch and local branch. The global branch is used to capture cross-scale long-range dependence to cope with large-scale defects, and the local branch is used to enhance the high-frequency texture of small defects in the tile. The dual-branch residual fusion outputs a calibration single-scale feature map. The weighted multi-scale fusion unit configures learnable weights for input features at each scale through cross-scale bidirectional skip connections and adopts a fast normalization fusion strategy to adaptively allocate the contribution of features at different resolutions to defect detection.

8. The high-precision defect detection method for ceramic tiles based on multi-scale feature refinement according to claim 7, characterized in that, In step S2, the processing flow of the BiFPN bidirectional feature pyramid network integrating the GLSA local self-attention mechanism includes: Receive the final feature map output by the backbone network through the CADown lightweight downsampling module; Each layer of the original final feature map is fed into an independent pre-feature calibration unit to complete feature calibration. The processing flow for a single final feature map is as follows: the final feature map is split into a global branch and a local branch for parallel processing; the global branch models long-range dependencies across pixels through global self-attention operations, extracts global semantic features of large-size defects, and obtains the global branch output features; the local branch focuses on local pixel regions of the feature map through local self-attention operations, extracts high-frequency texture details of small defects, and obtains the local branch output features; the global branch output features and the local branch output features are fused using residual addition to obtain the calibrated single-scale feature map; The calibration single-scale feature maps, after being calibrated by the pre-feature calibration unit, are input into the weighted multi-scale fusion unit to build a multi-feature fusion node with bidirectional cross-scale jump connections from top to bottom and from bottom to top. Independent learnable weights are configured for each multi-feature fusion node, and the range of values ​​of all learnable weights is constrained by the ReLU activation function. A fast normalization fusion strategy is used to weight and sum the multiple input features in the same node to complete the adaptive aggregation of defect features at different resolutions, and output a multi-scale feature map with multi-level fusion, as shown in formula (6): （6） in, Represents multi-scale feature maps, Indicates the sequence number of the calibration single-scale feature map. Indicates the first Learnable weights corresponding to each calibration single-scale feature map Used to maintain numerical stability Indicates the sequence number of the calibration single-scale feature map. Indicates the first Learnable weights corresponding to each calibration single-scale feature map This represents the calibration of a single-scale feature map.

9. The high-precision defect detection method for ceramic tiles based on multi-scale feature refinement according to claim 1, characterized in that, In step S1, the ratio of the training set, validation set, and test set is 8:1:

1.

10. The high-precision defect detection method for ceramic tiles based on multi-scale feature refinement according to claim 1, characterized in that, The defect categories include edge abnormalities, corner abnormalities, white spot blemishes, aperture blemishes, black spot blemishes, and light spot blemishes.

Images