A method and device for detecting the surface of a textured tile and a tile production line

Abstract

The application discloses a texture ceramic tile surface detection method, device and ceramic tile production line. The texture ceramic tile surface detection method comprises the following steps: acquiring an original image, preprocessing the original image to obtain a pretreated image. A multi-scale image pyramid is constructed based on the pretreated image, and texture features of each scale image in the multi-scale image pyramid are extracted respectively. A multi-scale texture feature matrix is generated based on the texture features of each scale image. The complexity score of the multi-scale texture feature matrix is calculated to obtain a texture complexity score value, and an adaptive mapping relationship of a smoothing parameter is constructed based on the texture complexity score value. The corresponding pretreated image is optimized based on the smoothing parameter to obtain a target enhanced image. The target enhanced image is sliced to obtain a plurality of slice images. The plurality of slice images are input into a preset image detection model one by one, and a surface defect detection result of a ceramic tile blank is output.

Description

Technical Field

[0001] This invention relates to the field of ceramic tile production lines and image recognition technology, and in particular to a method, apparatus and production line for detecting textured ceramic tile surfaces. Background Technology

[0002] Large-size ceramic tiles and slabs have gradually become mainstream products in the building ceramics market due to their superior appearance and physical properties. In actual production, the inspection of surface defects of semi-finished ceramic tile blanks after printing with large-format inkjet printers is one of the key steps in controlling product quality and determining product grade.

[0003] Currently, the detection of surface defects in semi-finished ceramic tile blanks mainly relies on manual visual inspection. This has significant limitations in actual production. The complex industrial environment requires quality inspectors to observe semi-finished ceramic tile blanks on the production line for extended periods and at high frequencies, easily leading to visual fatigue and distraction. Furthermore, manual inspection is inherently a subjective process lacking quantitative data support. As working hours increase, inspectors are prone to errors in judgment, resulting in a significant increase in both missed and false detection rates, making it impossible to guarantee the stability of inspection quality.

[0004] On the other hand, due to the large surface area of ​​large-sized products, quality inspectors cannot clearly see the entire brick surface simply by standing to the side of the conveyor line. Therefore, existing production lines typically construct an additional steel pedestrian bridge spanning the conveyor line, where personnel sit and conduct overhead inspections. This inspection layout not only significantly increases equipment investment and subsequent maintenance costs but also poses certain safety hazards due to personnel working above the equipment for extended periods. Summary of the Invention

[0005] To address the technical problems existing in the prior art, the present invention aims to provide a method, apparatus, and production line for detecting the surface of textured ceramic tiles. This method employs an automated vision solution that combines multi-scale texture feature extraction, adaptive image enhancement, and slice detection, replacing traditional manual visual inspection and enabling the detection of surface defects on large-sized textured ceramic tiles.

[0006] The objective of this invention is achieved through the following technical solution:

[0007] In a first aspect, this application provides a method for detecting the surface of textured ceramic tiles, applied to a ceramic tile production line. The ceramic tile production line includes a conveyor line and a camera assembly disposed on the conveyor line. The camera assembly is used to acquire images of the surface of ceramic tile blanks located on the conveyor line. The method for detecting the surface of textured ceramic tiles includes the following steps:

[0008] Acquire the original image, which is the image of the surface of the ceramic tile blank captured by the camera component;

[0009] The original image is preprocessed to obtain a preprocessed image;

[0010] A multi-scale image pyramid is constructed based on the preprocessed image, and texture features of each scale image in the multi-scale image pyramid are extracted respectively.

[0011] The texture features of images at different scales are aligned and stitched together to generate a multi-scale texture feature matrix;

[0012] The complexity score of the multi-scale texture feature matrix is ​​calculated to obtain the texture complexity score value, and an adaptive mapping relationship of smoothing parameters is constructed based on the texture complexity score value.

[0013] Based on the smoothing parameters, the corresponding preprocessed image is optimized to obtain the target enhanced image;

[0014] The target enhanced image is sliced ​​to obtain multiple slice images;

[0015] Multiple slice images are input one by one into a preset image detection model, and the surface defect detection results of the ceramic tile blank are output.

[0016] In conjunction with the first aspect, this application also provides a first specific implementation of the first aspect, which constructs a multi-scale image pyramid based on the preprocessed image, specifically including:

[0017] The preprocessed image is subjected to at least two levels of downsampling to generate multiple images of different resolutions;

[0018] The multiple images at different resolutions are combined with the original scale image to construct a multi-scale image pyramid;

[0019] The multi-scale image pyramid includes at least an original scale image, a first downsampled scale image, and a second downsampled scale image; and the scale ratio between each scale image is a fixed ratio.

[0020] In conjunction with the first aspect, this application also provides a second specific implementation of the first aspect, which extracts the texture features of images at each scale in the multi-scale image pyramid, specifically including:

[0021] Gradient calculation is performed on the images at each scale in the multi-scale image pyramid to obtain the corresponding gradient feature maps, and gradient statistical features are calculated based on the gradient feature maps.

[0022] A gray-level co-occurrence matrix is ​​constructed for each scale of the image, and texture statistical features are calculated based on the gray-level co-occurrence matrix.

[0023] The gradient statistical features include at least the mean gradient and the gradient standard deviation; the gray-level co-occurrence matrix features include at least one of entropy, contrast, energy, homogeneity, and correlation.

[0024] In conjunction with the first aspect, this application also provides a third specific implementation of the first aspect, which aligns and stitches the texture features of images at different scales to generate a multi-scale texture feature matrix, specifically including:

[0025] The texture features of images at various scales are normalized to obtain normalized texture features;

[0026] Position alignment is performed on the normalized texture features of images at each scale;

[0027] The normalized texture features of the aligned images at each scale are stitched together in a preset order to generate a multi-scale texture feature vector.

[0028] The multi-scale texture feature vectors are matrixed to obtain a multi-scale texture feature matrix.

[0029] In conjunction with the first aspect, this application also provides a fourth specific implementation of the first aspect, which involves calculating a complexity score for the multi-scale texture feature matrix to obtain a texture complexity score value, and constructing an adaptive mapping relationship for smoothing parameters based on the texture complexity score value, specifically including:

[0030] The multi-scale texture feature matrix is ​​standardized to obtain a standardized feature matrix;

[0031] Principal component analysis is performed on the standardized feature matrix to obtain the first complexity score.

[0032] The standardized feature matrix is ​​regressed and predicted based on the random forest regression model to obtain the second complexity score; wherein, the training label of the random forest regression model is the optimal smoothing parameter corresponding to the optimization process;

[0033] The first complexity score and the second complexity score are fused together to obtain the texture complexity score.

[0034] A mapping function for smoothing parameters is constructed based on the texture complexity score to obtain the smoothing parameter λ; wherein, the texture complexity score and the smoothing parameter λ are positively correlated.

[0035] In conjunction with the first aspect, this application also provides a fourth specific implementation of the first aspect, wherein the target enhancement image is sliced ​​to obtain multiple slice images, specifically including:

[0036] The target enhanced image is divided into regions based on a preset sliding window to obtain multiple image blocks;

[0037] Edge filling is performed on the multiple image blocks to obtain multiple slice images;

[0038] The preset sliding window performs a sliding scan on the target enhancement image according to a preset step size.

[0039] Secondly, this application provides a textured tile surface inspection device, applied in a tile production line, the tile production line including a conveyor line and a camera assembly disposed on the conveyor line; the camera assembly is used to acquire images of the surface of the tile blank located on the conveyor line; the textured tile surface inspection device includes:

[0040] An acquisition module is used to acquire an original image, which is an image of the surface of a ceramic tile blank captured by the camera component;

[0041] A preprocessing module is used to preprocess the original image to obtain a preprocessed image;

[0042] A construction module is used to construct a multi-scale image pyramid based on the preprocessed image, and extract the texture features of each scale image in the multi-scale image pyramid respectively;

[0043] The generation module is used to align and stitch together the texture features of images at different scales to generate a multi-scale texture feature matrix.

[0044] The calculation module is used to calculate the complexity score of the multi-scale texture feature matrix, obtain the texture complexity score value, and construct an adaptive mapping relationship of smoothing parameters based on the texture complexity score value.

[0045] The optimization module is used to perform optimization processing on the corresponding preprocessed image based on the smoothing parameters to obtain the target enhanced image;

[0046] The slicing module is used to slice the target enhanced image to obtain multiple slice images;

[0047] The detection module is used to input multiple slice images one by one into a preset image detection model and output the surface defect detection results of the ceramic tile blank.

[0048] Thirdly, this application provides a ceramic tile production line, including a control system, a conveyor line, and an inkjet printer, a blank storage machine, and a camera assembly disposed on the conveyor line; the camera assembly is located on the discharge side of the inkjet printer and is used to capture images of the surface of the ceramic tile blanks located on the conveyor line; the control system is connected to the conveyor line, the inkjet printer, the blank storage machine, and the camera assembly respectively.

[0049] The control system is configured to perform the following steps:

[0050] Acquire the original image, which is the image of the surface of the ceramic tile blank captured by the camera component;

[0051] The original image is preprocessed to obtain a preprocessed image;

[0052] A multi-scale image pyramid is constructed based on the preprocessed image, and texture features of each scale image in the multi-scale image pyramid are extracted respectively.

[0053] The texture features of images at different scales are aligned and stitched together to generate a multi-scale texture feature matrix;

[0054] The complexity score of the multi-scale texture feature matrix is ​​calculated to obtain the texture complexity score value, and an adaptive mapping relationship of smoothing parameters is constructed based on the texture complexity score value.

[0055] Based on the smoothing parameters, the corresponding preprocessed image is optimized to obtain the target enhanced image;

[0056] The target enhanced image is sliced ​​to obtain multiple slice images;

[0057] Multiple slice images are input one by one into a preset image detection model, and the surface defect detection results of the ceramic tile blank are output.

[0058] Based on the surface defect detection results, the conveyor line, inkjet printer, and blank storage machine are controlled.

[0059] In conjunction with the third aspect, this application also provides a first specific embodiment of the third aspect, wherein the billet storage machine includes:

[0060] A frame, wherein the frame is equipped with a lifting mechanism;

[0061] A roller conveyor is installed in the bottom space of the frame. The roller conveyor includes multiple conveying rollers arranged at intervals along the first direction, with conveying gaps formed between adjacent conveying rollers. The roller conveyor is located in the conveying line.

[0062] The billet storage rack includes multiple vertical frames spaced apart along the first direction. Each vertical frame is provided with multiple support rods spaced apart along the vertical direction. Each vertical frame is respectively set to correspond to the conveying gap of the roller conveyor.

[0063] Among them, the support rods of multiple vertical frames at the same horizontal height together form a storage layer, thereby forming multiple storage layers spaced apart in the vertical direction on the storage rack;

[0064] The lifting mechanism drives the storage rack to move it vertically. When the lifting mechanism drives the storage rack to descend and causes one of the storage layers to pass through the conveying plane of the roller conveyor from above, the ceramic tile blanks on that storage layer are transferred to the roller conveyor. When the lifting mechanism drives the storage rack to rise and causes one of the storage layers to pass through the conveying plane of the roller conveyor from below, the ceramic tile blanks on the roller conveyor are transferred to that storage layer.

[0065] In conjunction with the third aspect, this application also provides a second specific embodiment of the third aspect, wherein multiple billet storage machines are provided, including a first billet storage machine located upstream of the conveyor line and a second billet storage machine located downstream of the conveyor line; the first billet storage machine, the inkjet printer, and the second billet storage machine are arranged sequentially along the flow direction of the conveyor line;

[0066] Specifically, controlling the conveyor line, inkjet printer, and blank storage machine based on the surface defect detection results includes:

[0067] When the surface defect detection result is deemed unqualified, the inkjet printer is instructed to stop operating.

[0068] The lifting mechanism of the first blank storage machine drives the blank storage rack to rise and causes one of the blank storage layers to pass through the conveying plane of the roller conveyor from below, and the first ceramic tile blank entering the roller conveyor is transferred to the blank storage layer, so as to store the first ceramic tile blank in the first blank storage machine.

[0069] The lifting mechanism of the second blank storage machine drives the blank storage rack to descend and causes one of the blank storage layers to pass through the conveying plane of the roller conveyor from below. The second ceramic tile blank on this blank storage layer is transferred to the roller conveyor to release the second ceramic tile blank from the second blank storage machine into the conveying line.

[0070] Compared with existing technologies, the method, apparatus, and tile production line for detecting textured tile surfaces provided by this invention have at least the following beneficial effects:

[0071] This application provides a method for detecting textured tile surfaces, applied to a tile production line. The tile production line includes a conveyor line and a camera assembly installed on the conveyor line; the camera assembly is used to acquire images of the tile blank surface located on the conveyor line. The method for detecting textured tile surfaces includes the following steps: acquiring an original image, which is an image of the tile blank surface acquired by the camera assembly; preprocessing the original image to obtain a preprocessed image; constructing a multi-scale image pyramid based on the preprocessed image, and extracting texture features from each scale of the multi-scale image pyramid; aligning and stitching the texture features of each scale image to generate a multi-scale texture feature matrix; calculating the complexity score of the multi-scale texture feature matrix to obtain a texture complexity score value, and constructing an adaptive mapping relationship for smoothing parameters based on the texture complexity score value; performing optimization processing on the corresponding preprocessed image based on the smoothing parameters to obtain a target enhanced image; slicing the target enhanced image to obtain multiple slice images; and inputting the multiple slice images one by one into a preset image detection model to output the surface defect detection results of the tile blank.

[0072] This solution automatically acquires images of the tile surface by installing camera components on the tile conveyor line, replacing traditional manual inspection from above. This eliminates the maintenance costs of steel structure pedestrian bridges and removes the safety hazards of working at heights. Furthermore, it constructs a fully automated machine vision inspection process, replacing traditional manual visual inspection and eliminating subjective misjudgments and missed detections caused by visual fatigue. Based on this, a multi-scale image pyramid is used to align and stitch texture features at different resolutions, solving the problem of identifying small defects in complex texture backgrounds. Then, image enhancement parameters are adaptively adjusted based on texture complexity scores to achieve differentiated optimization. Finally, large-format images are sliced ​​and fed one slice at a time into a deep learning model for detection. This ensures comprehensive detection coverage of large-size tiles while improving inference speed and the detection rate of minute defects, achieving high-precision, high-efficiency, and stable and reliable automated online detection of surface defects in textured tiles. Attached Figure Description

[0073] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0074] Figure 1 A simplified diagram illustrating the composition of a ceramic tile production line according to an embodiment of the present invention;

[0075] Figure 2 This is a three-dimensional structural diagram of a ceramic tile storage machine provided in an embodiment of the present invention;

[0076] Figure 3 A front view of a ceramic tile storage machine provided in an embodiment of the present invention;

[0077] Figure 4 This is an assembly diagram of a ceramic tile storage machine provided in an embodiment of the present invention;

[0078] Figure 5 This is a three-dimensional structural diagram of the billet storage rack provided in an embodiment of the present invention;

[0079] Figure 6 A schematic flowchart of a method for detecting the surface of textured ceramic tiles provided in an embodiment of the present invention;

[0080] Figure 7 This is a schematic diagram illustrating the multi-scale extraction of textured tile features provided in an embodiment of the present invention.

[0081] In the picture:

[0082] 100 - Frame; 110 - Drive assembly; 120 - Slide rail assembly; 130 - First stop; 140 - Second stop;

[0083] 200 - Roller conveyor; 210 - First baffle plate; 220 - Second baffle plate;

[0084] 300-Bill storage rack; 310-Connecting rack; 320-Grid rack; 330-Vertical frame; 340-Support rod; 350-Bill storage layer. Detailed Implementation

[0085] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0086] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0087] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can also refer to the internal connection of two components; and they can refer to a wireless connection or a wired connection. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0088] Furthermore, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0089] Currently, the detection of surface defects in semi-finished ceramic tile blanks mainly relies on manual visual inspection. This has significant limitations in actual production. The complex industrial environment requires quality inspectors to observe semi-finished ceramic tile blanks on the production line for extended periods and at high frequencies, easily leading to visual fatigue and distraction. Furthermore, manual inspection is inherently a subjective process lacking quantitative data support. As working hours increase, inspectors are prone to errors in judgment, resulting in a significant increase in both missed and false detection rates, making it impossible to guarantee the stability of inspection quality.

[0090] On the other hand, due to the large surface area of ​​large-sized products, quality inspectors cannot clearly see the entire brick surface simply by standing to the side of the conveyor line. Therefore, existing production lines typically construct an additional steel pedestrian bridge spanning the conveyor line, where personnel sit and conduct overhead inspections. This inspection layout not only significantly increases equipment investment and subsequent maintenance costs but also poses certain safety hazards due to personnel working above the equipment for extended periods.

[0091] To address this, this application provides a novel solution that automatically captures images of the brick surface by installing camera components on the tile conveyor line, replacing traditional manual inspection from above. This eliminates the maintenance costs associated with steel structure pedestrian bridges and removes safety hazards associated with working at heights. Furthermore, it establishes a fully automated machine vision inspection process, replacing traditional manual visual inspection and eliminating subjective misjudgments and missed inspections caused by visual fatigue.

[0092] Based on this, a multi-scale image pyramid is used to align and stitch texture features at different resolutions, solving the problem of difficult identification of small target defects in complex texture backgrounds. Then, the image enhancement parameters are adaptively adjusted according to the texture complexity score to achieve differentiated optimization processing. Finally, the large-format image is sliced ​​and fed into the deep learning model for detection one by one. This ensures comprehensive detection coverage of large-size tiles while improving inference speed and the detection rate of small defects, achieving high-precision, high-efficiency, and stable and reliable automated online detection of surface defects of textured tiles.

[0093] Example 1

[0094] This application provides a ceramic tile production line. In the ceramic tile production process chain, the production line typically covers multiple core stages such as pressing, drying, inkjet printing, glazing, firing, edge grinding, and polishing. Correspondingly, the ceramic tile production line includes a continuous brick press, a drying kiln, an inkjet printer, glazing equipment, a firing kiln, edge grinding equipment, and polishing equipment arranged sequentially along the production flow. It also includes a conveyor line for connecting the various functional devices to transport the ceramic tile blanks.

[0095] Specifically, the aforementioned functional devices are connected sequentially via a conveyor line, enabling the ceramic tile blanks to continuously flow along a preset production route, achieving automatic connection and continuous production between various processes. For example, the conveyor line of this application can be formed by combining multiple roller conveyors, belt conveyors, or other conveying equipment suitable for ceramic tile transport. Taking roller conveyors as an example, multiple roller conveyors are sequentially spliced ​​along the ceramic tile flow direction, thereby forming a continuous conveying channel running through the entire production line to ensure stable transport of the ceramic tile blanks between various process devices.

[0096] It should be noted that the tile production line is an existing production line in this field, and will not be described in detail here.

[0097] like Figure 1 As shown, the tile production line of this application embodiment includes a control system, a conveyor line, and an inkjet printer, a blank storage machine, and a camera assembly installed on the conveyor line; the camera assembly is located on the discharge side of the inkjet printer and is used to collect images of the surface of the tile blanks located on the conveyor line; the control system is connected to the conveyor line, the inkjet printer, the blank storage machine, and the camera assembly respectively.

[0098] In this embodiment, the conveyor line is used to continuously transport the ceramic tile blanks, allowing them to pass through the inkjet printing process, the inspection process, and subsequent glazing and other ceramic tile production processes in sequence, thereby forming a continuous production process.

[0099] Furthermore, an inkjet printer is installed on the conveyor line to perform inkjet printing on the surface of the ceramic tile blanks on the conveyor line, thereby forming semi-finished ceramic tile blanks with preset textures. Inkjet printers are commercially available products and will not be discussed in detail here, such as the HPEJET TW2300 large-format inkjet printer.

[0100] In one implementation, a blank storage machine is located upstream of the inkjet printer to temporarily store and buffer the production cycle of the ceramic tile blanks conveyed from the drying kiln. This allows for the matching and decoupling of the production cycle between the drying process and the subsequent inkjet printing process, thereby improving the continuity and stability of the entire production line.

[0101] In another implementation, the blank storage machine is located downstream of the inkjet printer to temporarily store and buffer the cycle time of the printed ceramic tile blanks. This allows for matching and decoupling of the production cycle time between the inkjet printing process and the subsequent inspection process, thereby improving the continuity and stability of the entire production line.

[0102] In this embodiment, a camera component is disposed on the discharge side of the inkjet printer to acquire images of the surface of the ceramic tile blank located on the conveyor line. Specifically, the camera component is used to perform online visual imaging on the surface of the ceramic tile blank after inkjet printing is completed and it is output to the discharge end of the inkjet printer to obtain image data including surface texture, defect information, and inkjet quality information.

[0103] Furthermore, the camera assembly may include an industrial camera and a light source structure configured to work with it. The light source structure is used to provide uniform illumination to the surface of the ceramic tile blank, thereby reducing ambient light interference and improving the stability and consistency of image acquisition.

[0104] Understandably, once the inkjet printing process is complete, the pattern, texture, and inkjet quality on the surface of the tile blank are formed. Therefore, image acquisition at this location can promptly obtain information on the surface state after inkjet printing, enabling the detection and analysis of the inkjet quality, pattern integrity, and surface defects on the tile blank surface.

[0105] In this embodiment, the control system is electrically connected to the conveyor line, inkjet printer, blank storage machine, and camera assembly, respectively, for unified and coordinated control of these components. Specifically, the control system controls the conveying cycle of the ceramic blanks according to the operating status of the conveyor line and controls the printing action of the inkjet printer, ensuring that the inkjet process and the conveying process are synchronized; simultaneously, the control system controls the temporary storage and release of ceramic blanks according to the inventory status of the blank storage machine, thereby achieving production cycle adjustment and buffer control.

[0106] In addition, the control system is also used to receive image data collected by the camera assembly and analyze the surface condition of the tile blank based on the image data, so as to realize online detection and feedback control of inkjet printing quality. For example, when inkjet defects, abnormal textures, or tile blank abnormalities are detected on the surface, the control system can send control commands to various functional devices to achieve closed-loop quality control.

[0107] In one specific implementation, the conveyor line is composed of multiple roller conveyors spliced ​​together. Specifically, the multiple roller conveyors are arranged sequentially along a preset conveying direction and connected end to end to form a continuous conveying path, thereby achieving stable transmission of ceramic tile blanks between various processes.

[0108] Furthermore, each roller conveyor includes a frame and several conveying roller groups mounted on the frame. The conveying roller groups are arranged at intervals along the conveying direction and are driven by a drive assembly to rotate synchronously or in segments to achieve rolling conveying of ceramic tile blanks.

[0109] In one specific implementation, multiple billet storage machines are provided, including a first billet storage machine located upstream of the conveyor line and a second billet storage machine located downstream of the conveyor line. The first billet storage machine, the inkjet printer, and the second billet storage machine are arranged sequentially along the flow direction of the conveyor line. This forms a double buffer structure of pre-storage and post-storage on the tile production line.

[0110] The first blank storage machine is located upstream of the inkjet printer. It is used to pre-store and regulate the cycle of the ceramic tile blanks to be inkjet printed, ensuring a stable supply of material into the inkjet printer and acting as a buffer when the conveying cycle fluctuates. The second blank storage machine is located downstream of the inkjet printer. It is used to temporarily store and transfer the ceramic tile blanks after inkjet printing, ensuring a smooth connection and buffer between the inkjet printing process and subsequent glazing or other processes.

[0111] Furthermore, both the first and second blank storage machines can be configured with a storage structure having multiple layers or multiple workstations to realize the layered storage or multi-workstation cyclic conveying of ceramic blanks, thereby improving storage capacity and circulation efficiency.

[0112] By setting up a double-sided buffer structure with the upstream first blank storage machine and the downstream second blank storage machine, the inkjet printer is placed in a relatively stable feeding and discharging environment. This reduces the impact of conveyor cycle fluctuations on inkjet printing quality and inspection rhythm, improving the continuity and stability of the entire line operation. Furthermore, it allows for emergency response when the inkjet printer malfunctions.

[0113] Those skilled in the art can make equivalent substitutions or adjustments to the number, capacity, form, and location of the billet storage machine according to the production line cycle requirements and spatial layout, without departing from the protection scope of this application.

[0114] like Figure 2 As shown, in one specific embodiment, the billet storage machine includes a frame 100, a roller conveyor 200, and a billet storage rack 300. The frame 100 is equipped with a lifting mechanism. The roller conveyor 200 passes through the bottom space of the frame 100 and includes multiple conveying rollers arranged at intervals along a first direction, with conveying gaps formed between adjacent conveying rollers. The billet storage rack 300 includes multiple vertical frames 330 arranged at intervals along the first direction. Each vertical frame 330 is provided with multiple support rods 340 arranged at intervals along the vertical direction, and each vertical frame 330 is respectively set to correspond to the conveying gap of the roller conveyor 200.

[0115] In one embodiment of this application, the tile blank storage machine is mainly used for material buffering and turnover on a tile production line to coordinate the production rhythm of upstream and downstream processes and improve the continuous operation capability of the production line. The tile blank storage machine mainly includes a frame 100, a roller conveyor 200, and a blank storage rack 300.

[0116] Among them, the frame 100 serves as the basic load-bearing and support frame of the whole, and a lifting mechanism is installed on it. The lifting mechanism is connected to the storage rack 300 for transmission, and is used to provide power for the vertical movement of the storage rack 300, thereby realizing the layer-by-layer stacking and retrieval of materials.

[0117] Specifically, the roller conveyor 200 is installed in the bottom space of the frame 100. The roller conveyor 200 can serve as one of the conveying nodes in the tile production line, used to transport tile blanks from upstream to downstream. Simultaneously, the roller conveyor 200 is also responsible for smoothly guiding the tile blanks from upstream into the storage machine, or outputting the buffered tile blanks to the downstream process. For example, the roller conveyor 200 includes multiple conveying rollers arranged at intervals along a first direction (i.e., the material's forward conveying direction), forming a horizontal conveying channel and creating a conveying plane on the upper surface of the roller conveyor 200. Importantly, a pre-defined conveying gap is provided between adjacent conveying rollers, providing clearance for the storage rack 300.

[0118] In one embodiment, the roller conveyor 200 includes a conveyor body and support legs located at the bottom of the conveyor body, with multiple conveying rollers rotatably mounted on the conveyor body. The lower part of the conveyor body is completely open, and the height of the support legs is greater than the height of the billet storage rack 300, thereby accommodating a vertical frame 330 that passes through the conveying gap.

[0119] It should be noted that the roller conveyor 200 is an existing conveyor product in this field and is a very mature piece of equipment, so it will not be discussed further here.

[0120] To facilitate the smooth transfer of ceramic tiles in conjunction with the roller conveyor 200, the storage rack 300 includes multiple vertical frames 330 spaced apart along the first direction. Each vertical frame 330 is provided with multiple support rods 340 arranged at intervals along the vertical direction. The multiple support rods 340 together form a three-dimensional multi-layer space for supporting and storing ceramic tiles.

[0121] Furthermore, in terms of spatial layout, each vertical frame 330 is respectively configured to correspond one-to-one with different conveying gaps on the roller conveyor 200. It can be understood that during actual operation, after a single layer of tiles is conveyed by the roller conveyor 200 to a predetermined position at the bottom of the frame 100 and stops, the lifting mechanism drives the storage rack 300 to rise. At this time, the support rods 340 on the multiple vertical frames 330 smoothly pass through the conveying gaps between adjacent conveying rollers from bottom to top, steadily lifting the tiles located on the conveying rollers, completing the buffering action of a single layer of tiles. As subsequent tiles are continuously input, the lifting mechanism, in conjunction with the storage rack 300, performs step-by-step lifting, thereby realizing the continuous storage or unloading of multiple layers of tiles.

[0122] like Figure 3 As shown in this embodiment, multiple vertical frames 330 and support rods 340 at the same horizontal height collectively form a storage layer 350, thereby constituting multiple storage layers 350 spaced vertically on the storage rack 300. A lifting mechanism drives the storage rack 300 to move it vertically. When the lifting mechanism drives the storage rack 300 downwards, causing one storage layer 350 to pass over the conveying plane of the roller conveyor 200 from above, the ceramic tile blanks on that storage layer 350 are transferred to the roller conveyor 200. When the lifting mechanism drives the storage rack 300 upwards, causing one storage layer 350 to pass over the conveying plane of the roller conveyor 200 from below, the ceramic tile blanks on the roller conveyor 200 are transferred to that storage layer 350.

[0123] In this embodiment, to achieve efficient and three-dimensional buffering and retrieval of ceramic tile blanks, multiple support rods 340 at the same horizontal height on the vertical frames 330 work together to form a storage layer 350 for stably supporting the ceramic tile blanks. With this spatial layout, multiple support rods 340 on each vertical frame 330 form multiple storage layers 350 spaced vertically along the entire storage rack 300, thereby maximizing the use of vertical space and forming a high-density three-dimensional storage structure.

[0124] In the specific operation, the lifting mechanism is connected to the billet storage rack 300 and is configured to control and drive the billet storage rack 300 to move vertically, so as to cooperate with the roller conveyor 200 to complete the automated material storage and retrieval cycle. Its storage and retrieval interaction is as follows:

[0125] During the blank storage (feeding) process, the roller conveyor 200 first transports the blanks to be stored to the corresponding position and stops. At this time, the lifting mechanism drives the empty or half-loaded blank storage rack 300 to move upward. When one of the empty blank storage layers 350 passes through the conveying plane of the roller conveyor 200 from bottom to top, multiple support rods 340 of that layer simultaneously rise from the conveying gap between the conveying rollers, lifting the blanks that are resting on the conveying plane, thereby completing the transfer of the blanks from the roller conveyor 200 to the blank storage layer 350.

[0126] During the unloading (discharging) process, the lifting mechanism drives the storage rack 300 downwards. As the storage rack 300 descends, and one layer 350 carrying ceramic tile blanks passes through the horizontal conveying plane of the roller conveyor 200 from top to bottom, the conveying gaps between the support rods 340 and adjacent conveying rollers correspond vertically. This allows the support rods 340 to smoothly pass downwards through the conveying plane, while the ceramic tile blanks originally supported by the support rods 340 are smoothly intercepted and transferred to the conveying rollers of the roller conveyor 200. Subsequently, the support rods 340 continue to descend to avoid this obstruction, and the roller conveyor 200 starts to output the ceramic tile blanks outwards.

[0127] Through the aforementioned "non-interference, penetrating" relative lifting and transfer logic, this application achieves seamless material transfer between the storage layer 350 and the conveying plane. This process requires no additional translational manipulators or push-pull structures, resulting in smooth transfer actions and significantly shortening the storage and retrieval cycle, thereby improving the overall automation level and operational continuity of the production line in the buffering and transfer stages.

[0128] In a preferred embodiment, to reduce the number of vertical frames 330 and the overall weight of the billet storage rack 300, two or more conveying rollers can be spaced between two adjacent vertical frames 330. Reducing the number of vertical frames 330 directly reduces the overall weight of the billet storage rack 300. Since the billet storage rack 300 is a frequently moving load on the lifting mechanism, reducing its weight effectively reduces the inertia during lifting, resulting in smoother and more precise start-stop actions.

[0129] like Figure 4 As shown, in one specific embodiment, the frame 100 is a gantry frame structure, including four vertically arranged columns and a crossbeam connected to the top of the four columns; the lifting mechanism includes a drive assembly 110 and a slide rail assembly 120, the slide rail assembly 120 includes a guide rail vertically fixed on the column, and a slider disposed on the outside of the connecting frame 310 and slidingly engaged with the lifting guide rail.

[0130] In one specific implementation, four columns are symmetrically distributed in a matrix on a horizontal plane, collectively forming a vertical three-dimensional space for accommodating the storage rack 300. A crossbeam connects the tops of the four columns into a closed unit, thus constructing a high-rigidity, high-stability gantry-type load-bearing base. This frame structure can not only effectively bear the gravitational load of the multi-layer storage rack 300 and a large quantity of ceramic tile blanks, but also effectively resist the inertial impact and torsional stress caused by frequent starts and stops and step-by-step lifting.

[0131] To ensure the stable and precise operation of the billet storage rack 300 in the vertical channel, the lifting mechanism mainly includes a drive assembly 110 for providing lifting power, and four sets of slide rail assemblies 120 for providing precise stroke constraints and support.

[0132] Specifically, four sets of slide rail assemblies 120 are respectively installed on four columns. Each set of slide rail assemblies 120 includes a guide rail vertically fixed on the column and a slider disposed on the outside of the connecting frame 310 and slidingly engaged with the guide rail. The connecting frame 310 is fixedly connected to the overall outer structure of the billet storage rack 300, so that the billet storage rack 300 is suspended and confined within the internal space of the frame 100 through the connecting frame 310.

[0133] In one embodiment, the drive assembly 110 includes a drive motor mounted on the frame 100, and a chain drive mechanism driven by the drive motor and connected to the connecting frame 310. Specifically, the chain drive mechanism typically includes a sprocket driven by the drive motor and a drive chain wound around the sprocket, with the ends or specific nodes of the drive chain mechanically anchored to the connecting frame 310. The drive motor outputs rotational power, which drives the drive chain to retract or circulate through the sprocket, thereby pulling the connecting frame 310 and the billet storage rack 300 fixed thereto to move up and down along the vertical guide rail.

[0134] In another embodiment, the drive assembly 110 includes a lead screw pair vertically mounted on the frame 100 and a drive motor connected to the lead screw pair, the lead screw pair being fixedly connected to the connecting frame 310.

[0135] It should be noted that chain drive mechanisms and lead screw pairs are both conventional techniques in the field, and will not be described in detail here, as they are techniques that can be reasonably implemented by those skilled in the art based on the description in this application.

[0136] It should be noted that chain drive mechanisms and lead screw pairs are both conventional techniques in the field, and will not be described in detail here, as they are techniques that can be reasonably implemented by those skilled in the art based on the description in this application.

[0137] like Figure 2 and Figure 4As shown, in one specific embodiment, the frame 100 is provided with a plurality of first stops 130, which are positioned above the billet storage rack 300. The first stops 130 are capable of abutting against the billet storage rack 300 when it rises, thereby limiting the rise of the billet storage rack 300. Specifically, the plurality of first stops 130 are located in the top space of the frame 100, and their positions are set in the area above the billet storage rack 300 beyond its normal lifting stroke. Each first stop 130 can be connected to the side wall of a column via a mounting base, for example, located at the top end of a guide rail. The first stops 130 may be covered with a layer of cushioning and shock-absorbing material, such as polyurethane or rubber pads, to reduce impact vibration and structural damage during physical collisions.

[0138] like Figure 2 and Figure 4 As shown, in another specific embodiment, the frame 100 is provided with a plurality of second stops 140, which are disposed below the billet storage rack 300. The second stops 140 can abut against the billet storage rack 300 when it descends, thereby limiting the descent of the billet storage rack 300. When the second stops 140 abut against the billet storage rack 300, one of the billet storage layers 350 of the billet storage rack 300 is flush with the conveying plane of the roller conveyor 200.

[0139] To precisely limit the descent stroke of the storage rack 300 and ensure the safety of equipment and materials under extreme operating conditions, multiple second stops 140 are provided at the bottom of the frame 100. The second stops 140 are spatially positioned in the movable lower region of the storage rack 300. When the second stops 140 abut against the storage rack 300, a specific storage layer 350 on the storage rack 300 will be perfectly aligned with the conveying plane of the roller conveyor 200, i.e., they will be completely flush. This allows the ceramic tile blanks to be directly conveyed downstream by the roller conveyor 200 without being affected by the storage rack 300.

[0140] Specifically, multiple second stops 140 are disposed in the bottom space of the frame 100, and are positioned below the billet storage rack 300 outside its normal lifting stroke. Each second stop 140 can be connected to the side wall of the column via a mounting base, for example, located at the bottom end of the guide rail. The second stops 140 may be covered with a layer of cushioning and shock-absorbing material, such as polyurethane or rubber pads, to reduce impact vibration and structural damage in the event of a physical collision.

[0141] like Figure 5As shown, in one specific embodiment, the billet storage rack 300 includes a connecting frame 310, a grid frame 320, and multiple vertical frames 330. The connecting frame 310 is bolted to a lifting mechanism. The grid frame 320 is fixedly connected to the bottom of the connecting frame 310. The multiple vertical frames 330 are respectively welded to the grid frame 320, and the multiple vertical frames 330 are fixedly suspended below the grid frame 320 at equal intervals along a first direction.

[0142] Specifically, the connecting frame 310 serves as the physical connection between the billet storage rack 300 and the lifting mechanism. Its main function is to transmit lifting power and bear the overall load. In one embodiment, the connecting frame 310 is connected to the slider and drive assembly 110 of the lifting mechanism via bolts. Bolting allows for the separation of the billet storage rack 300 from the lifting mechanism, facilitating independent maintenance, repair, or replacement of either the lifting mechanism or the billet storage rack 300, thus avoiding the risk of overall scrapping associated with traditional welding methods.

[0143] Furthermore, the grid frame 320 is fixedly connected to the bottom of the connecting frame 310, serving as the load-bearing skeleton of the entire tile storage rack 300. The grid frame 320 provides a connection base for multiple vertical frames 330. The multiple vertical frames 330 are respectively welded to the lower end face of the grid frame 320. Through the welding process, each vertical frame 330 and the grid frame 320 form a solid whole, ensuring that each vertical frame 330 will not loosen or shake when fully loaded with tile blanks.

[0144] like Figure 5 As shown, in one specific embodiment, the connecting frame 310, the grid frame 320, and the vertical frame 330 are all frame structures welded from multiple square tubes. The vertical frame 330 is a rectangular frame structure, with multiple support rods 340 arranged at equal intervals along the vertical direction, and the two ends of each support rod 340 are welded to the opposite inner sidewalls of the vertical frame 330.

[0145] In one specific implementation, to further enhance the structural strength and rigidity of the billet storage rack 300, the connecting frame 310, the grid frame 320, and the vertical frame 330 all adopt a frame structure welded from multiple square tubes. Compared with plates or other profiles, square tubes have better bending and torsional resistance, providing high-rigidity skeletal support for the entire billet storage rack 300, while being relatively lighter in weight. On the other hand, the vertical frame 330 is designed as a rectangular frame structure, and the rectangular frame design enhances the structural strength of the frame itself.

[0146] like Figure 4 As shown, in a preferred embodiment, the roller conveyor 200 is provided with a first drive cylinder and a first baffle plate 210 connected to the output end of the first drive cylinder; the first baffle plate 210 is located on the side of the storage rack 300.

[0147] The first drive cylinder drives the first blank-blocking plate 210 to switch between the storage position and the stop position. When the first blank-blocking plate 210 is in the stop position, the first blank-blocking plate 210 extends from below the conveying gap to stop the ceramic tile blank located on the roller conveyor 200. When the first blank-blocking plate 210 is in the storage position, the first blank-blocking plate 210 is located below the conveying roller.

[0148] To ensure the alignment and positioning of the tile blanks before lifting and to prevent them from slipping beyond the predetermined lifting area due to inertia, the roller conveyor 200 is also equipped with a first drive cylinder and a first stop plate 210 that is drively connected to the output end of the first drive cylinder. The first stop plate 210 is located beside the storage rack 300, forming a stop mechanism that physically limits the movement of the tile blanks.

[0149] Specifically, the first drive cylinder, as a power actuator, is configured to drive the first blank holder 210 to reciprocate between a "stop position" and a "received position". The conveying gap formed between adjacent conveying rollers of the roller conveyor 200 is used as the physical channel for the movement of the first blank holder 210.

[0150] Specifically, the working state of the first baffle plate 210 is as follows:

[0151] When in the stop position: the first drive cylinder extends and pushes the first stop plate 210, causing it to pass through the conveying gap from below and eventually be positioned above the conveying plane. At this time, the first stop plate 210 forms a barrier that can accurately intercept the ceramic tile blanks conveyed by the roller conveyor 200, keeping them in the predetermined storage and retrieval position, and ensuring that the support rod 340 can lift the ceramic tile blanks when the subsequent storage rack 300 is vertically raised and lowered.

[0152] When in the storage position: the first drive cylinder retracts and pulls the first baffle plate 210 downwards until it is completely retracted into the space below the conveyor roller. At this time, the conveying plane is restored to a flat and unobstructed state, and the ceramic tile blank can continue to be conveyed to the next process along the roller conveyor 200 without interference.

[0153] like Figure 4 As shown, in a preferred embodiment, the roller conveyor 200 is provided with a second drive cylinder and a second baffle plate 220 connected to the output end of the second drive cylinder; the billet storage rack 300 is disposed between the first baffle plate 210 and the second baffle plate 220.

[0154] The second drive cylinder drives the second blank plate 220 to switch between the storage position and the stop position. When the second blank plate 220 is in the stop position, the second blank plate 220 extends from below the conveying gap to stop the ceramic tile blank located on the roller conveyor 200. When the second blank plate 220 is in the storage position, the second blank plate 220 is located below the conveying roller.

[0155] In a preferred embodiment, to further improve the positioning accuracy and flexibility of the ceramic tile blanks at the storage and retrieval station, the roller conveyor 200 is also equipped with a second drive cylinder and a second stop plate 220 connected to the output end of the second drive cylinder. The storage rack 300 is physically positioned between the first stop plate 210 and the second stop plate 220, and this symmetrical layout constitutes a two-way limiting and constraining system for the ceramic tile blanks.

[0156] Specifically, the second drive cylinder, as an independent power output element, is arranged parallel or symmetrically with the first drive cylinder to drive the second stop plate 220 to perform controlled vertical switching between the "receiving position" and the "stop position". The second stop plate 220 also moves through the conveying gap between adjacent conveying rollers of the roller conveyor 200 to perform the interception action. Its functional logic complements that of the first stop plate 210.

[0157] Coordination of the stopping state: When the second stop plate 220 is in the stopping position, it extends from below the conveying gap to above the conveying plane. On the one hand, the second stop plate 220 blocks subsequent tile blanks from entering, affecting the storage or unloading operation of the storage rack 300. On the other hand, through the coordinated operation of the first stop plate 210 and the second stop plate 220, the tile blanks located in the storage rack 300 can be stopped in both directions, thus sealing the tile blanks within the corresponding vertical lifting area of ​​the storage rack 300.

[0158] Flexibility in the storage state: When the second blank plate 220 is in the storage position, it is completely retracted under the conveyor roller, ensuring the flatness of the conveying channel and not interfering with the normal conveying of the ceramic tile blank, thus ensuring the versatility of the roller conveyor 200's operation.

[0159] In a preferred embodiment, the frame is also equipped with sensors for real-time monitoring of material position, equipment operating posture, and workflow status. The inclusion of these sensors establishes the closed-loop control logic of the billet storage machine, significantly improving automation and operational reliability.

[0160] Specifically, the sensors may include material detection sensors installed on the conveying path of the roller conveyor, and displacement monitoring sensors installed on the frame or lifting mechanism.

[0161] For example, the material detection sensors are preferably photoelectric sensors, symmetrically distributed in the working area of ​​the roller conveyor located below the storage rack. Their main function is to detect in real time whether the ceramic tile blanks have been accurately delivered to the target storage position.

[0162] For example, the displacement monitoring sensor uses an array of multi-point proximity switches. This array is mounted on the uprights of the frame and arranged vertically to achieve discrete positioning monitoring of the lifting height of the billet storage rack. Specifically, the multi-point proximity switch array includes multiple proximity switches arranged linearly at intervals in the height direction, and the installation height of each proximity switch corresponds to a preset billet storage layer position on the billet storage rack (i.e., the elevation corresponding to the support rod). Simultaneously, on the outside of the vertical frame fixedly connected to the billet storage rack, a sensing trigger element (such as a metal sensing block or a magnetic trigger block) matching the aforementioned proximity switches is provided.

[0163] In a preferred embodiment, multiple billet storage machines are provided, including a first billet storage machine located upstream of the conveyor line and a second billet storage machine located downstream of the conveyor line; the first billet storage machine, the inkjet printer, and the second billet storage machine are arranged sequentially along the flow direction of the conveyor line.

[0164] Specifically, the first blank storage machine, the inkjet printer, and the second blank storage machine are arranged sequentially along the flow direction of the conveyor line, thereby forming a front-end buffer node and a back-end buffer node for materials before and after the inkjet printing process, so that the ceramic tile blanks have the ability to temporarily store and adjust the cycle before entering the inkjet printing process and after completing the inkjet printing process.

[0165] The first blank storage machine is located upstream of the inkjet printer. It is used to pre-store and buffer the cycle of the ceramic tile blanks to be inkjet printed, so that the ceramic tile blanks entering the inkjet printer are in a continuous and stable feeding state, and play a buffering and regulating role when there are fluctuations or uneven feeding in the upstream conveying.

[0166] The second blank storage machine is located downstream of the inkjet printer. It is used to temporarily store and adjust the output of the ceramic tile blanks after inkjet printing, so as to decouple the cycle time between the inkjet printing process and the subsequent inspection, sorting or stacking processes, thereby avoiding the impact of fluctuations in the downstream process on the continuous operation of the inkjet printer.

[0167] Furthermore, by setting up a first blank storage machine and a second blank storage machine upstream and downstream of the inkjet printer respectively, the entire conveyor line forms a double buffer structure before and after the inkjet printing process, thereby improving the continuous operation capability and process matching flexibility of the production line and reducing the risk of downtime or congestion caused by single-point cycle fluctuations.

[0168] Those skilled in the art can make equivalent substitutions or optimizations to the capacity structure, number of storage layers, and connection method with the conveyor line of the first and second billet storage machines according to the actual production line layout requirements, without departing from the protection scope of this application.

[0169] Example 2

[0170] refer to Figure 6 This invention provides a flowchart illustrating a method for inspecting the surface of textured ceramic tiles in Embodiment 2. This invention offers a method for inspecting the surface of textured ceramic tiles in a ceramic tile production line. This method automatically acquires images of the tile surface by installing a camera assembly on the tile conveyor line, replacing traditional manual inspection from a height. Furthermore, it employs an automated vision scheme that utilizes multi-scale texture feature extraction, adaptive image enhancement, and slice detection to achieve high accuracy and efficiency in detecting surface defects in large-sized textured ceramic tiles / large slabs.

[0171] The textured tile surface detection method of the present invention is applied to a tile production line and can be executed by a control system. The control system can be implemented in hardware and / or software, specifically the control system of the aforementioned tile production line. The control system can be configured in the electronic equipment of the tile production line, such as a controller, edge computing device, etc.

[0172] The textured tile surface inspection method of the present invention is implemented based on a tile production line, specifically as described in Example 1. Figure 6 As shown, the surface inspection method for textured tiles is as follows:

[0173] Step S110: Obtain the original image, which is the image of the surface of the ceramic tile blank captured by the camera component.

[0174] In one specific embodiment, the original image is an image frame captured by the camera assembly on the surface of the ceramic tile during the conveying process. The original image can be an RGB color image. Furthermore, the camera assembly may include an industrial camera and may be used in conjunction with a ring light source or a strip light source for synchronous supplementary lighting to improve the imaging stability of the ceramic tile surface texture information.

[0175] In one possible implementation, the original image can be a high-resolution image, for example, 8192×6000 pixels or 4096×3500 pixels. Specifically, the original image resolution should be no less than 1000×1000 pixels to ensure clear details.

[0176] Step S120: Preprocess the original image to obtain a preprocessed image.

[0177] In one specific embodiment, the preprocessing includes, but is not limited to, at least one of the following processing operations: image denoising, illumination equalization, distortion correction, color normalization, and contrast enhancement. For example, Gaussian filtering, median filtering, or bilateral filtering can be used to denoise the original image to suppress random noise during the acquisition process.

[0178] Furthermore, image brightness distribution can be optimized using histogram equalization or adaptive histogram equalization (CLAHE) to mitigate the effects of local overexposure or underexposure caused by uneven illumination. In another possible implementation, perspective correction or lens distortion correction can be performed on the image based on camera calibration parameters to improve the geometric consistency of subsequent texture feature extraction.

[0179] In one possible implementation, preprocessing may include converting the color image to a grayscale image to highlight defect features on the tile surface. Alternatively, high-frequency device noise from the image acquisition process may be removed using a specific filtering algorithm.

[0180] Step S130: Construct a multi-scale image pyramid based on the preprocessed image, and extract the texture features of each scale image in the multi-scale image pyramid respectively.

[0181] In one possible implementation, such as Figure 7 As shown, to extract richer texture information, a Gaussian pyramid can be constructed from the preprocessed image to generate image layers at three scales. For example, the three scales may include:

[0182] Scale 1 (Original Scale): Focuses on detailed texture features and is suitable for capturing subtle variations in texture, such as small imperfections or small grain textures on the surface of a tile.

[0183] Scale 2 (half downsampling): Used to capture larger-scale texture variations, effectively capturing the general texture patterns on the tile surface.

[0184] Scale 3 (one-third downsampling): Primarily used to capture texture features of larger areas on the tile surface, suitable for detecting overall texture patterns.

[0185] Specifically, the preprocessed image can be downsampled layer by layer to construct an image hierarchy structure with different resolutions, so as to characterize the variation characteristics of the surface texture of the tile at different spatial scales.

[0186] In one possible implementation, texture features are extracted on the image at each scale, and the texture features may include, but are not limited to, at least one of the following: Local Binary Pattern (LBP) features; Gray-Level Co-occurrence Matrix (GLCM) statistical features; Gabor filter response features; and Histogram of Gradient Directions (HOG) features.

[0187] In one specific implementation, for any scale image in the multi-scale image pyramid, seven texture features are extracted respectively; the seven texture features include mean gradient, gradient standard deviation, entropy, contrast, energy, homogeneity and correlation.

[0188] Step S140: Align and stitch together the texture features of images at each scale to generate a multi-scale texture feature matrix.

[0189] In one implementation, since images at different scales differ in resolution and feature dimensions, it is necessary to perform uniform alignment processing on texture features at each scale. Specifically, interpolation methods or feature resampling methods can be used to map features at different scales to a unified dimensional space, for example, scale alignment can be achieved through bilinear interpolation or linear interpolation.

[0190] Furthermore, the aligned texture features at each scale are stitched together in a preset order to generate a multi-scale texture feature matrix. In one possible implementation, the stitching method can be stitching by channel dimension, unfolding stitching by spatial dimension, or concatenating stitching by feature vector.

[0191] Step S150: Calculate the complexity score of the multi-scale texture feature matrix to obtain the texture complexity score value, and construct an adaptive mapping relationship of smoothing parameters based on the texture complexity score value.

[0192] In one possible implementation, principal component analysis (PCA) can be used to calculate the complexity score for each sample. Specific implementation details are described below.

[0193] Step S160: Perform optimization processing on the corresponding preprocessed image based on the smoothing parameters to obtain the target enhanced image.

[0194] In one possible implementation, the optimization process can be the L0 gradient minimization algorithm. The smoothing intensity of the L0 gradient minimization algorithm is adaptively adjusted by quantizing the complexity of the textured tile pattern, thereby achieving the goal of preserving defects and smoothing the background texture in textured tiles containing defects.

[0195] It should be noted that in the detection of surface defects in textured tiles, the multi-scale characteristics of the pattern mean that the surface simultaneously contains both small-scale detailed textures and large-scale coarse textures. The L0 gradient minimization algorithm can smooth the image texture, achieving the goal of preserving defects and smoothing the background texture in textured tiles containing defects, making the defects more distinguishable from the background and easier to detect.

[0196] However, the traditional L0 gradient minimization algorithm cannot adaptively adjust the smoothing intensity according to the complexity of the image, making it difficult to directly apply this algorithm to automated inspection in industrial fields.

[0197] To address the aforementioned technical problems, this application quantifies the complexity of the textured tile pattern in step S150, using it as a prerequisite for implementing adaptive smoothing in the L0 gradient minimization algorithm. Specifically, when executing the L0 gradient minimization algorithm, a smoothing parameter is mapped based on the texture complexity score. In this way, for tile areas with simple textures, a smaller smoothing parameter is applied, for example, to prevent over-smoothing from causing the loss of subtle defects; for areas with complex background textures, a larger smoothing parameter is applied to strongly suppress redundant background noise interference.

[0198] Therefore, this application endows the L0 gradient minimization algorithm with dynamic adjustment capability through complexity scoring adaptation. The final output target enhancement image not only filters out complex background textures but also achieves high-fidelity preservation of surface defects, providing a high-quality image source for subsequent model detection.

[0199] Step S170: Slice the target enhanced image to obtain multiple slice images.

[0200] To prevent the loss of features of small target defects (such as small target defects that account for as much as 85.7% in the original dataset) due to direct scaling of large images, a sliding window is used to slice the image.

[0201] For example, the sliding window is set to a size of 640×640 and an overlap ratio of 0.25. The sliding window slices the image starting from the top left corner and moves gradually to the lower right in steps of 480; if the slice position exceeds the image size, the image is filled with edge fill.

[0202] Step S180: Input multiple slice images one by one into the preset image detection model, and output the surface defect detection results of the ceramic tile blank.

[0203] In one implementation, the output surface defect detection result may include inference data for the sliced ​​image, and may be further mapped to global defect information of the tile blank surface. For example, the surface defect detection result specifically includes the following four data dimensions: defect category label, defect location information, and detection confidence.

[0204] For example, regarding defect category labels, the model outputs the category index or name to which the currently detected area belongs. Based on actual tile production needs, the preset models can identify categories including, but not limited to: cracks, black spots / blemishes, edge chipping, and glaze loss. Each detected defect is assigned a discrete category label.

[0205] Regarding defect localization information, the model outputs the spatial location information of the defect target in the sliced ​​image, represented by the coordinates of the predicted bounding box. Since the detection process is based on sliding window slicing, the system automatically performs reverse mapping processing: combining the starting coordinates and offset of the sliding window, the local coordinates in the slice are converted into global physical coordinates in the original high-resolution image of the tile blank. This step ensures the accuracy of defect localization on the surface of the tile blank.

[0206] Regarding detection confidence, the model outputs a probability value (ranging from 0 to 1) for each detected defect, representing the model's certainty about the prediction result. The confidence level is derived by weighting the class probability with the regression accuracy of the predicted bounding box.

[0207] In one possible implementation, the preset image detection model is a lightweight model based on the YOLO11 architecture. To improve the detection accuracy and efficiency for complex tile backgrounds and small target defects, the preset image detection model adopts the following network structure:

[0208] Multi-feature fusion backbone network: During feature extraction, the Fusion module fuses feature maps of different scales. Specifically, the current feature map and multiple low-level feature maps are selected, and the low-level feature maps are downsampled and aligned through average pooling before being concatenated and feature extracted, achieving the interaction of semantic and detailed information.

[0209] Lightweight Downsampling Module (MDown): The MDown module replaces ordinary convolution during network downsampling. The MDown module divides the input feature map into two branches. One branch is passed through a convolution with a kernel of 3×3 and a stride of 2, and the other branch is passed through a max pooling layer and a 1×1 convolutional layer. Finally, the two branches are concatenated and fused by channel concatenation.

[0210] Ghost Feature Module: This module generates Ghost features to replace some traditional convolutional operations. By introducing linear operations such as depthwise separable convolution, it generates non-redundant feature maps, reducing the number of model parameters and computational cost while maintaining accuracy.

[0211] EMA Attention Mechanism: An efficient multi-scale attention (EMA) mechanism is introduced at the connection between the backbone network and the neck network. The EMA module extracts channel and spatial information through parallel branches, and cross-fuses channel and spatial features through cross-spatial learning to generate multi-scale feature representations, enabling the model to effectively focus on indistinguishable defect targets.

[0212] In one specific implementation, constructing a multi-scale image pyramid based on the preprocessed image includes the following steps:

[0213] Step S210: Perform at least two levels of downsampling on the preprocessed image to generate multiple images with different resolutions.

[0214] Specifically, to achieve a collaborative representation of global and local features of the tile surface texture, this embodiment employs a Gaussian pyramid construction method. First, the preprocessed image is denoted as I0 (original scale image), and then downsampled after passing an anti-aliasing filter. For example, a 2×2 window is used for downsampling, where:

[0215] First-level downsampling: The original scale image I0 is downsampled to obtain the first downsampled scale image I1, which has a resolution of 1 / 2 that of the original image.

[0216] Second-level downsampling: The image I1 at the first downsampling scale is further downsampled to obtain the image I2 at the second downsampling scale, which has a resolution of 1 / 4 of the original image (i.e., a scale ratio of 1:4 relative to I0).

[0217] Step S220: Combine the multiple images of different resolutions with the original scale image to construct a multi-scale image pyramid.

[0218] The multi-scale image pyramid includes at least an original scale image, a first downsampled scale image, and a second downsampled scale image; and the scale ratio between each scale image is a fixed ratio.

[0219] Through the above processing, a pyramid model with a three-layer structure is constructed: The scale ratio between images at each scale is a fixed ratio (i.e., the ratio between adjacent layers is 2:1).

[0220] After constructing a multi-scale image pyramid, feature extraction for images at different scales has different focuses:

[0221] Original scale image: This layer is primarily used to extract texture abrupt changes caused by edges, cracks, and minute color differences, as it preserves pixel-level details. By applying GLCM calculations with a 3×3 neighborhood window to I0, it is possible to accurately capture the fine pinhole or crack textures generated during the ceramic firing process.

[0222] First downsampling scale image: This scale is mainly used to extract the periodic patterns of texture, such as the layout of printed patterns on the surface of tiles. By performing gradient intensity statistics on this level, some high-frequency device noise can be effectively filtered out, while preserving the directional features of the main texture.

[0223] Second downsampling scale image: This scale is used to extract the macroscopic color patch distribution and overall color uniformity features of the tile surface. Due to the reduced image resolution, this level can avoid interference from excessively fine local textures, thus providing a more stable global complexity assessment.

[0224] By spatially aligning and stitching the texture feature matrices extracted from these three scales, a multi-scale texture feature matrix is ​​constructed that can comprehensively reflect the all-round surface state of the tile from "local defects" to "macroscopic patterns," providing robust data support for subsequent complexity scoring and adaptive mapping of smoothing parameters.

[0225] In one specific implementation, the texture features of images at each scale in the multi-scale image pyramid are extracted, specifically including the following steps:

[0226] Step S310: Perform gradient calculation on each scale image in the multi-scale image pyramid to obtain the corresponding gradient feature map, and calculate gradient statistical features based on the gradient feature map.

[0227] Specifically, to quantify the roughness and edge sharpness of the tile surface, this embodiment applies the Sobel operator to perform convolution operations on images at various scales to obtain gradient information in the horizontal and vertical directions, thereby constructing gradient feature maps. Based on the gradient feature maps, the mean gradient and gradient standard deviation are statistically calculated to characterize the overall undulation and the degree of local variation in the tile texture.

[0228] For example, the mean gradient feature measures the magnitude of brightness variation in a local region of an image, reflecting the coarseness of the texture. The formula for calculating the mean gradient is as follows:

[0229] In the formula, This represents the pixel grayscale value of the image at coordinates (x, y); The horizontal and vertical gradients are calculated for the Sobel operator, where N is the total number of pixels.

[0230] in, The horizontal gradient value obtained by applying the Sobel operator to coordinates (x, y). The Sobel operator is also used to detect horizontal edges. The vertical gradient value obtained by applying the Sobel operator to the coordinates (x,y).

[0231] The standard deviation of the gradient characterizes the dispersion of the gradient distribution, describing the dispersion of local brightness changes in an image, and is a measure of texture detail changes. A larger standard deviation of the gradient indicates more drastic texture changes. The formula for calculating the standard deviation of the gradient is as follows:

[0232] In the formula, The gradient magnitude for all pixels; Calculate the difference (deviation) between the gradient magnitude of each pixel and the average gradient magnitude of the entire image.

[0233] Step S320: Construct gray-level co-occurrence matrices for images at each scale, and calculate texture statistical features based on the gray-level co-occurrence matrices.

[0234] The gradient statistical features include at least the mean gradient and the gradient standard deviation; the gray-level co-occurrence matrix features include at least one of entropy, contrast, energy, homogeneity, and correlation.

[0235] To further describe the spatial distribution patterns on the tile surface, this embodiment employs the gray-level co-occurrence matrix method. First, the gray levels of the image at each scale are quantized (e.g., compressed from 256 levels to 16 or 32 levels to reduce computational complexity). Then, the gray-level co-occurrence probabilities of adjacent pixel pairs are statistically analyzed at specific directions (typically 0°, 45°, 90°, and 135°) and distances (d=1), generating a matrix. .

[0236] Based on the gray-level co-occurrence matrix, the following five core texture statistical features are calculated:

[0237] Entropy: Reflects the complexity and disorder of a texture. For example, based on the Gray-Level Co-occurrence Matrix (GLCM), the formula for calculating entropy is: ;in, This represents the probability that gray levels i and j co-occur given a direction and distance.

[0238] Contrast ratio: Contrast ratio is the range of grayscale values ​​in an image, reflecting the changes in brightness and darkness of the image texture. High contrast textures exhibit obvious light and dark contrast, while low contrast textures appear smoother. .

[0239] Energy: The energy feature measures the concentration of gray values ​​in an image and is usually related to the uniformity of texture. A high energy value indicates a more uniform image texture, while low energy means a more complex texture. .

[0240] Homogeneity: Homogeneity measures the similarity of pixels within an image region. High homogeneity indicates that the texture within the region is relatively consistent, while low homogeneity indicates that the texture varies significantly. .

[0241] Correlation: Correlation describes the linear relationship between pixels in an image and reflects the structural information in a textured image. High correlation indicates a regular texture structure, while low correlation indicates that the texture has greater variation. .in, The row and column mean of GLCM The standard deviation is denoted as .

[0242] Through the collaborative processing of steps S810 and S820, this embodiment can extract a 7-dimensional feature vector containing 2 gradient features and 5 GLCM features from each scale image. Combined with the previously constructed three-scale image pyramid (Scale1, Scale2, Scale3), a 21-dimensional comprehensive texture feature matrix is ​​finally formed. This not only utilizes gradient features to capture high-frequency details such as edges and cracks, but also uses GLCM statistical features to characterize the spatial correlation of macroscopic patterns on the tile surface, thus providing a complete and multi-dimensional data foundation for the adaptive mapping of complexity scoring and smoothing parameters in subsequent steps.

[0243] In one possible implementation, after calculating the seven types of features mentioned above at each scale, a multi-scale feature matrix F is formed. This feature matrix can be represented as:

[0244] Where n is the number of samples, and the 21-dimensional feature consists of 3 scales multiplied by 7 features.

[0245] In one specific implementation, the texture features of images at different scales are aligned and stitched together to generate a multi-scale texture feature matrix, which specifically includes the following steps:

[0246] Step S410: Normalize the texture features of images at each scale to obtain normalized texture features.

[0247] Because texture features extracted at different scales (original scale, first downsampling scale, and second downsampling scale) have different dimensions and numerical ranges (for example, the value of gradient features is usually much larger than the value of entropy features), direct concatenation will result in uneven numerical distribution, leading to deviations in subsequent complexity scoring calculations. This embodiment uses Z-score Normalization to uniformly process the seven extracted texture features.

[0248] For feature data at each scale Normalization is performed using the following formula. ;in, This is the mean of the feature across all samples. This represents the standard deviation of the feature. This step maps all feature values ​​to a standard normal distribution interval with a mean of 0 and a standard deviation of 1, eliminating the interference of dimensional differences on model evaluation.

[0249] Step S420: Align the normalized texture features of images at each scale.

[0250] Because images at different scales have different spatial resolutions, the spatial coordinates of the original scale image are used as a reference when performing position alignment. Bilinear interpolation or a mapping function is used to ensure that the texture feature points extracted from the downsampled scale image are semantically aligned with the texture feature points of the original scale image. This step ensures that the stitched feature vector can accurately map back to the physical location on the surface of the ceramic tile, avoiding spatial drift of the features.

[0251] Step S430: The normalized texture features of the aligned images at each scale are spliced ​​together in a preset order to generate a multi-scale texture feature vector.

[0252] To ensure consistency of feature inputs, this embodiment employs a fixed, preset order for concatenation. Specifically, the construction of feature vectors follows a "hierarchical priority" principle:

[0253] First, arrange the 7 texture features of the original scale image (Scale1). Then, arrange the 7 texture features of the first downsampled scale image (Scale2). Finally, arrange the 7 texture features of the second downsampled scale image (Scale3). The generated feature vector V is represented as: This results in a feature vector with a total dimension of 21. This vector fully covers the full-scale feature information of the tile surface, from microscopic details to macroscopic patterns.

[0254] Step S440: Perform matrix processing on the multi-scale texture feature vector to obtain a multi-scale texture feature matrix.

[0255] The generated 21-dimensional feature vector is integrated with multiple tile sample data continuously collected on the production line to form a multi-scale texture feature matrix. , where N is the number of samples in the current processing batch, and 21 is the feature dimension.

[0256] Through the systematic processing of steps S410 to S440 described above, this application effectively transforms unstructured texture information in the multi-scale spatial domain into structured matrix data. This matrix format can not only be directly used as input for subsequent random forest models or complexity scoring models, but also preserves the correlation between features at different scales. This allows for a balance between the sharpness of local details and the smoothness of the global distribution when evaluating the complexity of tile surfaces, significantly improving the accuracy of subsequent defect detection.

[0257] In some possible implementations, in order to quantify the complexity of the tile surface texture and provide the optimal model selection basis for subsequent defect detection tasks, this embodiment constructs independent texture complexity scoring models based on the extracted multi-scale multi-feature dataset through principal component analysis, kernel principal component analysis, and random forest, respectively.

[0258] In one specific implementation, a complexity score is calculated on the multi-scale texture feature matrix to obtain a texture complexity score value, and an adaptive mapping relationship for smoothing parameters is constructed based on the texture complexity score value, specifically including the following steps:

[0259] Step S510: Standardize the multi-scale texture feature matrix to obtain a standardized feature matrix.

[0260] To eliminate bias caused by different dimensions of features, mean-variance standardization is used to process the multi-scale texture feature matrix. The mean of each feature in the matrix is ​​subtracted and then divided by the standard deviation, mapping all features to a distribution interval with a mean of 0 and a variance of 1.

[0261] Step S520: Perform principal component analysis on the standardized feature matrix to obtain the first complexity score.

[0262] In one possible implementation, Principal Component Analysis (PCA) is used to calculate the first complexity score for each sample. The core idea of ​​PCA is to map the data from the original feature space to a new feature space through orthogonal transformation, thereby maximizing the variance of the data in the new space.

[0263] Specifically, the complexity of the texture is quantified by considering the variance contribution of each principal component and its projection onto the sample. This dimensionality reduction process not only removes redundant features but also effectively reflects the variability of the texture.

[0264] It should be noted that after determining the principal components, a comprehensive score needs to be calculated for the image features. The eigenvalue of each principal component represents the degree of contribution of that component to the total variance of the data. The complexity score is calculated by combining the variance contribution of each principal component.

[0265] When calculating the complexity score, PCA focuses on the contribution of these principal components. Larger eigenvalues ​​mean that the principal component can better represent the variation in data variance, thus reflecting the complexity of the image.

[0266] For example, the specific process for calculating the first complexity score can be derived using the following formula: First, calculate the projection values ​​of the image features along each principal component direction. The projection values ​​of the image features along each principal component direction are obtained by calculating the inner product between the feature vector of the image and the feature vector of the principal component:

[0267] ;in, Let m be the projection of the image onto the principal component m. The feature vector of the image, It is the eigenvector of the m-th principal component.

[0268] Subsequently, the overall complexity score (i.e., the first complexity score) is obtained by multiplying the projection value of each feature of the image onto the principal components with the cumulative feature value of each principal component:

[0269] ;in, is the eigenvalue of principal component m, and k is the number of principal components to be retained. To ensure the interpretability of the data, the top k principal components with a cumulative variance contribution rate of 90% can be retained for the above calculation.

[0270] Step S530: Perform regression prediction on the standardized feature matrix based on the random forest regression model to obtain the second complexity score; wherein, the training label of the random forest regression model is the optimal smoothing parameter corresponding to the optimization process.

[0271] In one specific implementation, this step employs a Random Forest Regression model, leveraging its ensemble learning capabilities to process high-dimensional data. In this model, the training labels are set to the optimal smoothing parameters corresponding to the L0 gradient minimization optimization process. Through training, the Random Forest model learns the intrinsic relationship between feature distribution and optimal smoothing strength. The second complexity score, S2, represents the model's regression output.

[0272] Specifically, the process of calculating the complexity score using the random forest regression model includes the following implementation details:

[0273] First, the standardized multi-scale texture feature matrix (e.g., a 21-dimensional feature vector covering 3 scales and 7 features) is used as the input features of the model. The training labels of the random forest regression model are set to the optimal smoothing parameters corresponding to the optimization process (e.g., L0 gradient minimization algorithm). Through this supervised training method, the model can directly learn the complex nonlinear mapping relationship between multi-scale spatial domain features and the optimal smoothing intensity, making the output score applicable to practical applications.

[0274] During the training phase, the Random Forest algorithm randomly selects samples and features from the training set to construct multiple regression decision trees. Each decision tree is trained on a random subset of the training data. Each tree node selects the optimal feature from the chosen feature subset for splitting, until a preset stopping condition is met. During the inference phase, the average of the regression results output by all decision trees is used as the second complexity score for the current input sample.

[0275] In a preferred implementation, to further balance the model's high prediction accuracy with the risk of overfitting, the hyperparameters of the random forest regression model are set as follows:

[0276] (1) Number of trees: Set to 200. Random forests make predictions by ensemble of multiple decision trees. Increasing the number of trees can cover most data patterns and make the model more robust.

[0277] (2) Number of features used in each split: set to the square root of the total number of features. Since the selected texture features include features that are positively and negatively correlated with complexity, using a feature subset at the square root level can reduce the correlation between trees, increase the randomness of the model, and thus improve the generalization ability.

[0278] (3) Minimum number of leaf node samples: set to 5. This parameter controls the minimum number of samples required for each leaf node; a moderate limit on the number of leaf node samples will force the model to generate trees of appropriate depth, avoiding overfitting of the training data due to overly complex trees.

[0279] Through the above step S530, this embodiment utilizes the advantages of ensemble learning to achieve high-precision regression prediction of the complexity of tile texture. The obtained second complexity score can accurately reflect the optimal smoothness level required for the current texture image.

[0280] Step S540: The first complexity score and the second complexity score are fused to obtain the texture complexity score.

[0281] In one specific implementation, a weighted fusion strategy is used to calculate the final texture complexity score. :

[0282] ;in, The preset weighting coefficients, and By integrating the data, the robustness of complexity assessment is significantly improved, and the abnormal fluctuations that may be caused by a single model are reduced.

[0283] Step S550: Construct a mapping function for smoothing parameters based on the texture complexity score to obtain the smoothing parameter λ; wherein, the texture complexity score and the smoothing parameter λ are positively correlated.

[0284] In one specific implementation, a texture complexity score is established. With smoothing parameters A positive correlation mapping relationship exists between them. As texture complexity increases, smoothing parameters are adjusted to better preserve defects. The value needs to be increased accordingly. The mapping function can be defined as a linear or sigmoid function:

[0285] ;

[0286] in, This is the positive gain coefficient. This is the basic smoothing bias term. Through this function, the system can calculate the optimal smoothing intensity in real time based on the texture features of the current tile blank, achieving adaptive control of the image preprocessing process and ensuring the feature extraction efficiency of the target enhanced image in subsequent defect detection.

[0287] In some implementations, the coefficient and Instead of being a preset fixed value, it is calculated using the Ordinary Least Squares (OLS) fitting technique based on the "complexity score - optimal smoothing parameter" benchmark dataset in the sample library.

[0288] In one specific implementation, the target enhancement image is sliced ​​to obtain multiple slice images, specifically including the following steps:

[0289] Step S610: Divide the target enhancement image into regions based on a preset sliding window to obtain multiple image blocks.

[0290] Specifically, considering that the original images collected from the tile production line have high-resolution features (e.g., 8192×6000), directly scaling them and inputting them into the detection model would cause the features of small-sized defect targets to be smoothed or lost during downsampling. Therefore, this embodiment uses a sliding window scanning mechanism to divide the target enhancement image into regions.

[0291] Furthermore, the size of the sliding window is set to... (e.g., 640×640 pixels) The sliding window performs a sliding scan on the target enhancement image from left to right and from top to bottom according to a preset stride (e.g., 480 pixels). Since the stride is smaller than the window size, there is a preset overlap area between adjacent sliding windows (e.g., overlap rate of 0.25). The design of this overlap area is intended to ensure that the defect target located at the edge of the sliding window can be completely included in at least one slice image, avoiding feature truncation caused by the slice boundary effect, thereby improving the recall rate of small target defects.

[0292] Step S620: Edge filling is performed on the multiple image blocks to obtain multiple slice images. The preset sliding window performs a sliding scan on the target enhanced image according to a preset step size.

[0293] In practice, when the sliding scan reaches the right or bottom boundary of the target augmented image, if the size of the remaining image area is smaller than the size of the preset sliding window, it cannot directly form a complete training or inference data block. Therefore, this embodiment performs edge padding to ensure that all sliced ​​images have a uniform size specification.

[0294] Furthermore, the specific methods of edge filling include, but are not limited to, one of the following:

[0295] Zero padding: Fills the area outside the image area with pixels with a gray value of 0.

[0296] Replication Padding: Copies the pixels at the image boundary and expands them to meet the window size.

[0297] Reflection Padding: The image content is mirrored with the boundary as the axis of symmetry, and then filled into the missing area.

[0298] Through the above filling process, multiple slice images with uniform specifications are obtained, thereby satisfying the hard constraint of the preset image detection model (i.e., the lightweight model based on YOLO11) on the fixed input size (e.g., 640×640 pixels) in the subsequent steps, and ensuring the stability of the detection model when processing ceramic tile blank images of different sizes.

[0299] In one specific implementation, the control system is further configured, based on the above, to execute the following control method: controlling the conveyor line, inkjet printer, and blank storage machine according to the surface defect detection results.

[0300] Specifically, this application constructs an integrated closed-loop system of detection, decision-making, and feedback, achieving a high degree of automation in the production line.

[0301] Specifically, based on the surface defect detection results, controlling the conveyor line, inkjet printer, and blank storage machine includes:

[0302] Step S710: When the surface defect detection result is characterized as unqualified, the inkjet printer is instructed to stop operation.

[0303] Specifically, if the tile blank currently passing through the inspection area is determined to be a defective product (such as damaged tile blank, printing abnormalities, or missing texture), the control system sends a stop command to the inkjet printer.

[0304] Step S720: Control the lifting mechanism of the first blank storage machine to drive the blank storage rack to rise and cause one of the blank storage layers to pass through the conveying plane of the roller conveyor from below, and transfer the first ceramic tile blank into the roller conveyor to the blank storage layer, so as to store the first ceramic tile blank in the first blank storage machine.

[0305] Step S730: Control the lifting mechanism of the second blank storage machine to drive the blank storage rack to descend and cause one of the blank storage layers to pass through the conveying plane of the roller conveyor from below, and the second ceramic tile blank on the blank storage layer is transferred to the roller conveyor to release the second ceramic tile blank of the second blank storage machine into the conveying line.

[0306] This application also provides a textured tile surface inspection device, applied to the control system of the tile production line described in Embodiment 1, to execute and implement a textured tile surface inspection method as described in Embodiment 2. Specifically, the textured tile surface inspection device includes:

[0307] The acquisition module is used to acquire the original image, which is the image of the surface of the ceramic tile blank captured by the camera component;

[0308] The preprocessing module is used to preprocess the original image to obtain a preprocessed image;

[0309] The module is used to construct a multi-scale image pyramid based on the preprocessed image and extract the texture features of each scale image in the multi-scale image pyramid.

[0310] The generation module is used to align and stitch together the texture features of images at different scales to generate a multi-scale texture feature matrix.

[0311] The calculation module is used to calculate the complexity score of the multi-scale texture feature matrix, obtain the texture complexity score value, and construct an adaptive mapping relationship of smoothing parameters based on the texture complexity score value.

[0312] The optimization module is used to perform optimization processing on the corresponding preprocessed image based on the smoothing parameters to obtain the target enhanced image;

[0313] The slicing module is used to slice the target enhancement image to obtain multiple slice images;

[0314] The detection module is used to input multiple slice images one by one into a preset image detection model and output the surface defect detection results of the ceramic tile blank.

[0315] The above embodiments are merely preferred embodiments of the present invention and should not be construed as limiting the scope of protection of the present invention. For those skilled in the art, it will be understood that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for inspecting the surface of textured ceramic tiles, applied in a ceramic tile production line, characterized in that, The tile production line includes a conveyor line and a camera assembly installed on the conveyor line; the camera assembly is used to acquire images of the surface of the tile blanks located on the conveyor line; the textured tile surface detection method includes the following steps: Acquire the original image, which is the image of the surface of the ceramic tile blank captured by the camera component; The original image is preprocessed to obtain a preprocessed image; A multi-scale image pyramid is constructed based on the preprocessed image, and texture features of each scale image in the multi-scale image pyramid are extracted respectively. The texture features of images at different scales are aligned and stitched together to generate a multi-scale texture feature matrix; The complexity score of the multi-scale texture feature matrix is ​​calculated to obtain the texture complexity score value, and an adaptive mapping relationship of smoothing parameters is constructed based on the texture complexity score value. Based on the smoothing parameters, the corresponding preprocessed image is optimized to obtain the target enhanced image; The target enhanced image is sliced ​​to obtain multiple slice images; Multiple slice images are input one by one into a preset image detection model, and the surface defect detection results of the ceramic tile blank are output.

2. The method for detecting the surface of textured ceramic tiles according to claim 1, characterized in that, Constructing a multi-scale image pyramid based on the preprocessed image specifically includes: The preprocessed image is subjected to at least two levels of downsampling to generate multiple images of different resolutions; The multiple images at different resolutions are combined with the original scale image to construct a multi-scale image pyramid; The multi-scale image pyramid includes at least an original scale image, a first downsampled scale image, and a second downsampled scale image; and the scale ratio between each scale image is a fixed ratio.

3. The method for detecting the surface of textured ceramic tiles according to claim 2, characterized in that, Extracting texture features from images at each scale in the multi-scale image pyramid, specifically including: Gradient calculation is performed on the images at each scale in the multi-scale image pyramid to obtain the corresponding gradient feature maps, and gradient statistical features are calculated based on the gradient feature maps. A gray-level co-occurrence matrix is ​​constructed for each scale of the image, and texture statistical features are calculated based on the gray-level co-occurrence matrix. The gradient statistical features include at least the mean gradient and the gradient standard deviation; the gray-level co-occurrence matrix features include at least one of entropy, contrast, energy, homogeneity, and correlation.

4. The method for detecting the surface of textured ceramic tiles according to claim 3, characterized in that, The texture features of images at different scales are aligned and stitched together to generate a multi-scale texture feature matrix, specifically including: The texture features of images at various scales are normalized to obtain normalized texture features; Position alignment is performed on the normalized texture features of images at each scale; The normalized texture features of the aligned images at each scale are stitched together in a preset order to generate a multi-scale texture feature vector. The multi-scale texture feature vectors are matrixed to obtain a multi-scale texture feature matrix.

5. The method for detecting the surface of textured ceramic tiles according to claim 1, characterized in that, The complexity score of the multi-scale texture feature matrix is ​​calculated to obtain a texture complexity score value, and an adaptive mapping relationship of smoothing parameters is constructed based on the texture complexity score value, specifically including: The multi-scale texture feature matrix is ​​standardized to obtain a standardized feature matrix; Principal component analysis is performed on the standardized feature matrix to obtain the first complexity score. The standardized feature matrix is ​​regressed and predicted based on the random forest regression model to obtain the second complexity score; wherein, the training label of the random forest regression model is the optimal smoothing parameter corresponding to the optimization process; The first complexity score and the second complexity score are fused together to obtain the texture complexity score. A mapping function for smoothing parameters is constructed based on the texture complexity score to obtain the smoothing parameter λ; wherein, the texture complexity score and the smoothing parameter λ are positively correlated.

6. The method for detecting the surface of textured ceramic tiles according to claim 1, characterized in that, The target enhancement image is sliced ​​to obtain multiple slice images, specifically including: The target enhanced image is divided into regions based on a preset sliding window to obtain multiple image blocks; Edge filling is performed on the multiple image blocks to obtain multiple slice images; The preset sliding window performs a sliding scan on the target enhancement image according to a preset step size.

7. A textured tile surface inspection device, applied in a tile production line, characterized in that, The tile production line includes a conveyor line and a camera assembly installed on the conveyor line; the camera assembly is used to capture images of the surface of the tile blanks located on the conveyor line; the textured tile surface detection device includes: An acquisition module is used to acquire an original image, which is an image of the surface of a ceramic tile blank captured by the camera component; A preprocessing module is used to preprocess the original image to obtain a preprocessed image; A construction module is used to construct a multi-scale image pyramid based on the preprocessed image, and extract the texture features of each scale image in the multi-scale image pyramid respectively; The generation module is used to align and stitch together the texture features of images at different scales to generate a multi-scale texture feature matrix. The calculation module is used to calculate the complexity score of the multi-scale texture feature matrix, obtain the texture complexity score value, and construct an adaptive mapping relationship of smoothing parameters based on the texture complexity score value. The optimization module is used to perform optimization processing on the corresponding preprocessed image based on the smoothing parameters to obtain the target enhanced image; The slicing module is used to slice the target enhanced image to obtain multiple slice images; The detection module is used to input multiple slice images one by one into a preset image detection model and output the surface defect detection results of the ceramic tile blank.

8. A ceramic tile production line, characterized in that, The system includes a control system, a conveyor line, and an inkjet printer, a blank storage machine, and a camera assembly installed on the conveyor line. The camera assembly is located on the discharge side of the inkjet printer and is used to capture images of the surface of the ceramic tile blanks on the conveyor line. The control system is connected to the conveyor line, the inkjet printer, the blank storage machine, and the camera assembly. The control system is configured to perform the following steps: Acquire the original image, which is the image of the surface of the ceramic tile blank captured by the camera component; The original image is preprocessed to obtain a preprocessed image; A multi-scale image pyramid is constructed based on the preprocessed image, and texture features of each scale image in the multi-scale image pyramid are extracted respectively. The texture features of images at different scales are aligned and stitched together to generate a multi-scale texture feature matrix; The complexity score of the multi-scale texture feature matrix is ​​calculated to obtain the texture complexity score value, and an adaptive mapping relationship of smoothing parameters is constructed based on the texture complexity score value. Based on the smoothing parameters, the corresponding preprocessed image is optimized to obtain the target enhanced image; The target enhanced image is sliced ​​to obtain multiple slice images; Multiple slice images are input one by one into a preset image detection model, and the surface defect detection results of the ceramic tile blank are output. Based on the surface defect detection results, the conveyor line, inkjet printer, and blank storage machine are controlled.

9. A ceramic tile production line according to claim 8, characterized in that, The billet storage machine includes: A frame, wherein the frame is equipped with a lifting mechanism; A roller conveyor is installed in the bottom space of the frame. The roller conveyor includes multiple conveying rollers arranged at intervals along the first direction, with conveying gaps formed between adjacent conveying rollers. The roller conveyor is located in the conveying line. The billet storage rack includes multiple vertical frames spaced apart along the first direction. Each vertical frame is provided with multiple support rods spaced apart along the vertical direction. Each vertical frame is respectively set to correspond to the conveying gap of the roller conveyor. Among them, the support rods of multiple vertical frames at the same horizontal height together form a storage layer, thereby forming multiple storage layers spaced apart in the vertical direction on the storage rack; The lifting mechanism drives the storage rack to move it vertically. When the lifting mechanism drives the storage rack to descend and causes one of the storage layers to pass through the conveying plane of the roller conveyor from above, the ceramic tile blanks on that storage layer are transferred to the roller conveyor. When the lifting mechanism drives the storage rack to rise and causes one of the storage layers to pass through the conveying plane of the roller conveyor from below, the ceramic tile blanks on the roller conveyor are transferred to that storage layer.

10. A ceramic tile production line according to claim 9, characterized in that, The billet storage machine is provided in multiple ways, including a first billet storage machine located upstream of the conveyor line and a second billet storage machine located downstream of the conveyor line; the first billet storage machine, the inkjet printer and the second billet storage machine are arranged sequentially along the flow direction of the conveyor line; Specifically, controlling the conveyor line, inkjet printer, and blank storage machine based on the surface defect detection results includes: When the surface defect detection result is deemed unqualified, the inkjet printer is instructed to stop operating. The lifting mechanism of the first blank storage machine drives the blank storage rack to rise and causes one of the blank storage layers to pass through the conveying plane of the roller conveyor from below, and the first ceramic tile blank entering the roller conveyor is transferred to the blank storage layer to store the first ceramic tile blank in the first blank storage machine. The lifting mechanism of the second blank storage machine drives the blank storage rack to descend and causes one of the blank storage layers to pass through the conveying plane of the roller conveyor from below. The second ceramic tile blank on this blank storage layer is transferred to the roller conveyor to release the second ceramic tile blank from the second blank storage machine into the conveying line.

Images