CN117115095A

CN117115095A - Method and system for detecting tiny defects of ceramic tiles with complex textures

Info

Publication number: CN117115095A
Application number: CN202310997780.0A
Authority: CN
Inventors: 陈新度; 邱伟彬; 吴磊; 甘胜斯
Original assignee: Guangdong University of Technology
Current assignee: Guangdong University of Technology
Priority date: 2023-08-08
Filing date: 2023-08-08
Publication date: 2023-11-24

Abstract

The application relates to a method for detecting small defects of a ceramic tile with complex textures, which comprises the following steps: acquiring a tile image to be detected; preprocessing the tile image to be detected to obtain a first image; performing defect detection on the first image through a pretrained improved small target scale YOLOv8 network model to obtain a defect detection result; outputting the defect detection result to finish defect detection; specifically, the improved small target scale YOLOv8 network model performs defect positioning of texture tile images based on MAE mask self-coding network, the positioned defect information is further fused into the YOLOv8 network to be trained, and a small target detection layer is added, wherein the small target detection layer is used for enhancing the detection capability of small targets of the YOLOv8 network model. The application can solve the problems that the prior art only has good effect on the middle and large defects of plain tiles, but the detection effect is not ideal for tiles with complex textures and fine defects in the tiles.

Description

Method and system for detecting tiny defects of ceramic tiles with complex textures

Technical Field

The application relates to the technical field of machine vision, in particular to a method and a system for detecting small defects of a ceramic tile with complex textures.

Background

Conventional visual inspection techniques for tile surface defects include both manual-based visual inspection and computer-vision-based automatic inspection. Manual-based visual inspection requires a lot of manpower and time, and it is difficult to ensure the accuracy of the inspection. The existing method for automatically detecting the tile defects by using machine vision only has good effect on the middle and large defects of plain tiles and has certain limitation, and the method particularly shows the following two aspects:

firstly, aiming at small defects in ceramic tiles, namely small defects (common defect types have pinholes, black spot impurities and the like) with the size ranging from 0.5 mm to 2mm, the problem of missed detection often occurs.

Secondly, for the ceramic tile with complex texture, the interference of the complex texture background can cause the problem of false detection of the model. Meanwhile, if a defect happens to be in the texture, the defect target may be ignored, so that missed detection is possible.

Disclosure of Invention

The application aims to at least solve one of the defects in the prior art and provides a method and a system for detecting small defects of a ceramic tile with complex textures.

In order to achieve the above purpose, the present application adopts the following technical scheme:

specifically, a method for detecting small defects of a ceramic tile with complex textures is provided, which comprises the following steps:

acquiring a tile image to be detected;

preprocessing the tile image to be detected to obtain a first image;

performing defect detection on the first image through a pretrained improved small target scale YOLOv8 network model to obtain a defect detection result;

outputting the defect detection result to finish defect detection;

specifically, the improved small target scale YOLOv8 network model performs defect positioning of texture tile images based on MAE mask self-coding network, the positioned defect information is further fused into the YOLOv8 network to be trained, and a small target detection layer is added, wherein the small target detection layer is used for enhancing the detection capability of small targets of the YOLOv8 network model.

Further, specifically, the tile image to be detected is preprocessed to obtain a first image, which includes,

and (3) performing angular point positioning on the tile image to be detected, removing useless background environments, extracting tile objects, and performing filtering denoising treatment to obtain a first image.

Further, in particular, the process of pre-training an improved fine-scale YOLOv8 network model, includes,

collecting images of a plurality of related complex texture tiles;

preprocessing the acquired image to obtain a data set of the complex texture tile, wherein the data set comprises a defect-free data set and a defect data set;

training the MAE network by using the defect-free data set to obtain MAE network weights;

using a defect data set, and obtaining defect positioning information through the trained MAE network;

improving a characteristic pyramid structure in the YOLOv8 network, and adding a fine target detection layer to obtain an improved YOLOv8 network;

and converting the defect positioning information into a binarization channel, integrating the binarization channel with the defect picture, inputting the binarization channel into an improved YOLOv8 network, training the network weight of the improved YOLOv8 network, and further obtaining an improved small target scale YOLOv8 network model.

Further, the method also comprises the steps of,

after preprocessing the collected image, image segmentation processing is also performed, and the collected image is segmented into small pictures with preset specifications, so that the training speed of the model is improved.

Further, in particular, the process of training the MAE network, includes,

randomly selecting part of non-defective pictures in the non-defective data set, inputting the non-defective pictures into an encoder of an MAE network, and performing random mask processing on the input original pictures according to a preset rule and in a proportion of 75% to obtain processed pictures;

in the decoder, pixels of the mask portion of the processed picture are restored, and then MSE between the predicted pixels and the original pixels is calculated as Loss, so as to obtain and store a trained MAE weight file.

Further, specifically, the feature pyramid structure in the YOLOv8 network is improved, and the optimization process of adding a fine target detection layer comprises,

extracting upper layer characteristics with target information on the basis of an original YOLOv8 network on the first C2f layer of a backhaul;

the highest-level features of the FPN pyramid are up-sampled again, fused with the upper-level features and input into a PAN network together;

and inputting the first layer characteristics of the PAN network into the Head to obtain a fine target detection layer.

The application also provides a system for detecting the tiny defects of the ceramic tile with the complex texture, which comprises:

the conveyor belt is used for conveying the ceramic tiles to be detected;

the photoelectric sensor is used for judging whether the ceramic tile transmitted by the conveyor belt reaches a preset position or not;

the image acquisition device comprises a color line scanning camera and a linear light source, wherein the linear light source is used for irradiating the conveyor belt, and the color line scanning camera is used for shooting a large-format super-resolution image of the ceramic tile;

the PLC device is electrically connected with the photoelectric sensor and the controller of the conveyor belt and is used for acquiring a trigger signal of the photoelectric sensor and a speed parameter of the controller;

the PC is electrically connected with the image acquisition device and the PLC device and is used for receiving the data transmitted by the PLC device and receiving the image information acquired by the image acquisition device, the PC comprises,

the data acquisition module is used for acquiring the tile image to be detected;

the preprocessing module is used for preprocessing the tile image to be detected to obtain a first image;

the defect detection module is used for carrying out defect detection on the first image through a pretrained improved small target scale YOLOv8 network model to obtain a defect detection result;

the result output module is used for outputting the defect detection result and finishing defect detection;

The application also proposes a computer readable storage medium storing a computer program which, when executed by a processor, implements the steps of a method for implementing fine defect detection of a complex texture tile as described.

The beneficial effects of the application are as follows:

the application provides a method for detecting small defects of a complex texture tile, which is convenient for collecting texture tile data sets required by network training and completing the defect detection work of the tile by constructing an industrial detection line system platform.

Secondly, by optimizing the YOLOv8 network and utilizing the shallow layer characteristics extracted from the main network, a fine target detection layer is specifically added, so that the model has the detection capability of fine targets.

In addition, the self-supervision MAE method is used for completing defect positioning of the texture tile image, defect information is further fused into a YOLOv8 network, additional space information is provided, defect characteristics are enhanced, interference of texture background is reduced, defect detection efficiency of a model on the complex texture tile is improved, and the model has generalization and robustness.

Drawings

The above and other features of the present disclosure will become more apparent from the detailed description of the embodiments illustrated in the accompanying drawings, in which like reference numerals designate like or similar elements, and which, as will be apparent to those of ordinary skill in the art, are merely some examples of the present disclosure, from which other drawings may be made without inventive effort, wherein:

FIG. 1 is a flow chart of a method for detecting fine defects of a ceramic tile with complex textures according to the present application;

FIG. 2 is a schematic diagram showing an implementation of a method for detecting fine defects of a ceramic tile with complex textures according to the present application;

FIG. 3 is a schematic hardware diagram of a system for detecting fine defects of a ceramic tile with complex textures according to the present application;

FIG. 4 is a diagram showing the effect of MAE mask self-encoding network for implementing the method for detecting fine defects of ceramic tiles with complex textures;

FIG. 5 is a schematic diagram of the structure of an improved fine target scale YOLOv8 network model for implementing the method for detecting fine defects of a complex texture tile according to the present application;

fig. 6 is a comparison chart of detection results after reference of a method for detecting fine defects of a ceramic tile with complex textures.

Detailed Description

The conception, specific structure, and technical effects produced by the present application will be clearly and completely described below with reference to the embodiments and the drawings to fully understand the objects, aspects, and effects of the present application. It should be noted that, without conflict, the embodiments of the present application and features of the embodiments may be combined with each other. The same reference numbers will be used throughout the drawings to refer to the same or like parts.

Referring to fig. 1, embodiment 1 of the present application provides a method for detecting fine defects of a complex texture tile, comprising the following steps:

step 110, acquiring a tile image to be detected;

step 120, preprocessing the tile image to be detected to obtain a first image;

130, performing defect detection on the first image through a pretrained improved small target scale YOLOv8 network model to obtain a defect detection result;

step 140, outputting the defect detection result to finish defect detection;

The industrial detection line system platform is built so as to collect texture tile data sets required by network training and finish the defect detection work of the tiles.

Secondly, by optimizing the YOLOv8 network, a fine target detection layer is specifically added, so that the model has the detection capability of fine targets.

In addition, the self-supervision MAE method is used for completing defect positioning of the texture tile image, defect information is further fused into a YOLOv8 network, the defect characteristics are enhanced, the interference of texture background is reduced, the defect detection efficiency of the model on the complex texture tile is improved, and the model has generalization and robustness.

MAE is a self-supervising method that randomly masks the input image and then reconstructs the missing pixels. The MAE consists of an asymmetric encoder-decoder, where the encoder operates on only a subset of the data that is not masked, and the decoder can reconstruct the original image from the underlying representation and the mask. Since the defects do not belong to the original tile texture layout, the MAE reconstructed image will not restore the defective areas. And comparing the pixel difference between the reconstructed image and the original image to finish the positioning of the defect. The defect positioning information is utilized to be arranged into binary channel information, the binary channel information is input into a pretreatment network of YOLO, additional target space information is provided, meanwhile, the characteristics of a defect area are enhanced, the interference of texture background can be reduced, and the defect detection accuracy is improved.

The method is based on the fact that tile defect detection is achieved by using a deep learning method, self-supervision is achieved by using a mask self-encoder, defect target positioning is achieved, a feature pyramid structure in a YOLOv8 network is optimized, a feature detection layer with a small scale is added for small target defects, and the detection capability of the network for the small target defects is enhanced. After training of the complex texture tile is completed, the complex texture tile is deployed on an actual industrial real-time detection line, so that the accuracy of the detection line on the detection of the small defects of the complex texture tile is improved.

The feature map in the convolutional neural network has a shallow layer and a deep layer, wherein the feature map extracted by the shallow layer network has smaller receptive field, and the information is the features related to the pixels, such as color, texture, position and the like, and lacks guidance of advanced semantic information. The deep network extracts high-level semantic information such as the shape, outline, parts and the like of an object, the extracted feature map is small, and the range of the local receptive field on the original image is large, namely the deep network can acquire the global information of the original image, which is not beneficial to extracting the feature information of a small target. In the original YOLOv8 network, a PAFPN (Feature Pyramid Network, path Aggregation Network) structure is used, and shallow and deep feature maps are fused, where the structure has 3 detection layers with different scales for detecting large, medium and small targets, and the minimum range of the original map perceived is 8x8 pixel area, and if the defect is wider or higher than 8 pixels, it is possible to cause missed detection. Therefore, to further improve the accuracy of detecting the fine target, a shallow layer feature is extracted on the backbone network backhaul, and a fine target detection layer is added, so that the network can sense the pixel area with the original image range accurate to 4x4, as shown in fig. 5.

As a preferred embodiment of the present application, specifically, the pre-processing is performed on the tile image to be detected to obtain a first image, including,

As a preferred embodiment of the present application, specifically, the process of pre-trained improved fine-scale YOLOv8 network model, includes,

collecting images of a plurality of related complex texture tiles;

As a preferred embodiment of the application, the method further comprises,

As a preferred embodiment of the present application, specifically, the process of training the MAE network, includes,

As a preferred embodiment of the present application, in particular, the feature pyramid structure in YOLOv8 network is improved, and the optimization process of adding a fine target detection layer includes,

the conveyor belt is used for conveying the ceramic tiles to be detected;

In particular, the implementation flow of the application is shown in figure 2,

1. on a hardware platform as shown in fig. 3, acquiring images of related complex texture tiles;

2. performing image processing, enhancement and labeling on the acquired image to obtain a data set of the complex texture tile;

3. training the MAE network by using the defect-free data set to obtain MAE network weights;

4. using the defect data set, obtaining defect positioning information through MAE, as shown in FIG. 4, wherein the leftmost side is a mask image, the middle is a reconstructed image, and the rightmost side is an original image;

5. improving the feature pyramid structure in the YOLOv8 network, and adding a fine target detection layer (a dotted line box part in fig. 5);

6. converting the defect positioning information into a binarization channel, inputting the binarization channel and the defect picture into an improved YOLO v8 network, and completing training to obtain the network weight of YOLO;

7. and deploying the trained YOLOv8 network weight on an actual industrial detection line to complete real-time detection of the small defects of the complex texture tile.

In detail, the application generally comprises the following parts,

1. constructing a dataset

1.1 building a tile image acquisition platform, as shown in fig. 2. The platform consists of a color line scanning camera, a linear light source, a photoelectric sensor, a PLC, a PC, a conveyor belt and a ceramic tile object to be detected. The conveyor belt is responsible for completing the horizontal movement of the tiles, and the photoelectric sensor is responsible for detecting whether tiles come. The PLC is responsible for transmitting the sensor signals and the conveyor speed parameters to the computer. The color line scanning camera can completely shoot a large-format super-resolution image of the ceramic tile under the action of the linear light source. The PC is responsible for receiving data and coordinating the work among the hardware.

1.2, acquiring texture tile pictures by using the built camera platform, positioning the corner points of the tile objects, removing useless background environments, extracting the tile objects, and then filtering and denoising.

1.3 the tile object resolution obtained is 8192 x 8192, whereas the defect object occupies only a small part of its area, so to increase the training speed of the model, the picture is divided into a plurality of small pictures with resolution 512 x 512.

And 1.4, screening out defective objects for marking, and obtaining a defect data set of the texture tile. Meanwhile, the tiles without defective objects are reserved.

2. Training MAE mask self-encoding network

2.1 randomly selecting a small part of the non-defective pictures, inputting the pictures into an encoder of the MAE, and carrying out random masking on the original pictures according to a certain rule and a proportion of 75 percent.

2.2 in the decoder, the masking part of pixels are restored, MSE between the predicted pixels and the original pixels is calculated as loss, and the trained MAE weight file is saved for preparation of the next step.

3. Optimizing YOLOv8-PAFPN networks

3.1 on the basis of the original YOLOv8 network, extracting the upper layer characteristics with target information from the first C2f layer of the backsheeded.

And 3.2, upsampling the highest level features of the FPN pyramid, fusing the highest level features with the above-mentioned backstone upper level features, and inputting the fused features into a PAN network.

And 3.3, inputting the first layer of characteristics of the PAN network into the Head, and obtaining a detection layer with a small target scale. When the input image size is 512×512, the feature map size of the obtained small target scale detection layer is 128×128, and the feature map with the size has smaller receptive field, so that smaller targets can be detected.

4. Training YOLOv8 networks

4.1 using the defect dataset described above, defects were localized by MAE and converted to binary information.

4.2 training using the modified YOLO model with defect localization binarized information as additional channels to input the modified YOLO model.

4.3, training is completed, and a weight file with detection capability on small defects of the ceramic tile with complex texture is obtained

And 4.4, deploying the YOLO weight file on an operation computer of an actual industrial line, namely processing the complex texture ceramic tile transmitted by the color line scanning camera in real time, and finishing detection of defects in the complex texture ceramic tile. Fig. 6 shows a specific example effect, and it can be seen that the present application has a good detection effect when applied.

The modules described as separate components may or may not be physically separate, and components shown as modules may or may not be physical modules, i.e., may be located in one place, or may be distributed over a plurality of network modules. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution in this embodiment.

In addition, each functional module in each embodiment of the present application may be integrated into one processing module, or each module may exist alone physically, or two or more modules may be integrated into one module. The integrated modules may be implemented in hardware or in software functional modules.

The integrated modules, if implemented in the form of software functional modules and sold or used as a stand-alone product, may be stored in a computer readable storage medium. Based on this understanding, the present application may implement all or part of the flow of the method of the above embodiment, or may be implemented by a computer program to instruct related hardware, where the computer program may be stored in a computer readable storage medium, and when the computer program is executed by a processor, the computer program may implement the steps of each of the method embodiments described above. Wherein the computer program comprises computer program code which may be in source code form, object code form, executable file or some intermediate form etc. The computer readable medium may include: any entity or system capable of carrying the computer program code, a recording medium, a U disk, a removable hard disk, a magnetic disk, an optical disk, a computer Memory, a Read-Only Memory (ROM), a random access Memory (RAM, random Access Memory), an electrical carrier signal, a telecommunications signal, a software distribution medium, and so forth.

While the present application has been described in considerable detail and with particularity with respect to several described embodiments, it is not intended to be limited to any such detail or embodiments or any particular embodiment, but is to be construed as providing broad interpretation of such claims by reference to the appended claims in view of the prior art so as to effectively encompass the intended scope of the application. Furthermore, the foregoing description of the application has been presented in its embodiments contemplated by the inventors for the purpose of providing a useful description, and for the purposes of providing a non-essential modification of the application that may not be presently contemplated, may represent an equivalent modification of the application.

The present application is not limited to the above embodiments, but is merely preferred embodiments of the present application, and the present application should be construed as being limited to the above embodiments as long as the technical effects of the present application are achieved by the same means. Various modifications and variations are possible in the technical solution and/or in the embodiments within the scope of the application.

Claims

1. The method for detecting the fine defects of the ceramic tile with the complex texture is characterized by comprising the following steps of:

acquiring a tile image to be detected;

preprocessing the tile image to be detected to obtain a first image;

outputting the defect detection result to finish defect detection;

2. The method for detecting fine defects of ceramic tiles with complex textures according to claim 1, wherein the step of preprocessing the image of the ceramic tile to be detected to obtain a first image comprises,

3. A method for performing fine defect detection on complex texture tiles according to claim 1, wherein the process of the pre-trained modified fine-scale Yolov8 network model comprises,

collecting images of a plurality of related complex texture tiles;

4. A method for performing fine defect detection on a complex texture tile according to claim 3, further comprising,

5. A method for performing fine defect detection on complex texture tiles according to claim 3, wherein the training of the MAE network comprises,

6. A method for performing fine defect detection on a complex-textured tile according to claim 3, wherein the optimization process for improving the feature pyramid structure in the YOLOv8 network comprises,

7. A system for implementing fine defect detection of a complex texture tile, comprising:

the conveyor belt is used for conveying the ceramic tiles to be detected;

8. A computer readable storage medium storing a computer program, characterized in that the computer program when executed by a processor implements the steps of the method according to any of claims 1-6.