A defect image processing method for autoclaved lightweight new wall material

Abstract

The present invention proposes a method for processing defect images of autoclaved lightweight new wall materials, and relates to the field of image enhancement. The present invention comprises the steps of collecting a plurality of defect images of autoclaved lightweight new wall materials, constructing a feature transformation module, constructing a feature adaptive processing module, constructing a feature enhancement module, constructing a feature optimization module, constructing a defect image processing model of autoclaved lightweight new wall materials, and training and processing the defect image processing model of autoclaved lightweight new wall materials. Meanwhile, a defect image processing model of autoclaved lightweight new wall materials is proposed, wherein the feature transformation module improves the feature expression capability, the feature adaptive processing improves the feature matching degree and the processing effect, and the feature optimization module further improves the feature recognition degree and the processing effect through the operation of the channel dimension.

Description

Technical Field

[0001] The invention belongs to the field of image enhancement, and in particular relates to a defect image processing method of an autoclaved lightweight novel wall material. Background Art

[0002] Autoclaved lightweight new wall materials are widely used in construction projects and have good thermal insulation, heat preservation, fire resistance and other properties. However, due to the influence of production process, environmental conditions and construction quality, this type of material may have defects such as cracks, voids and deformation during use. These defects not only affect the overall performance of the material, but may also lead to building safety hazards, so the detection and treatment of material defects is crucial.

[0003] Image processing technology provides an effective means for detecting material defects. By processing the defect images of wall materials, the defect features can be displayed more clearly, which helps to improve the accuracy and efficiency of detection. However, since defect images are often affected by problems such as uneven lighting and noise interference, it is often difficult to achieve ideal results by directly performing defect detection. Therefore, before performing defect detection, it is usually necessary to enhance the image to improve the image quality and make the defect features more prominent.

[0004] Image enhancement technology improves the visual effect of images and highlights key features by adjusting image contrast, brightness, sharpening, etc. In the processing of defect images of autoclaved lightweight new wall materials, commonly used image enhancement methods include histogram equalization, adaptive filtering, edge enhancement, etc. These technologies can effectively remove noise from images and improve image clarity, providing a more accurate basis for subsequent defect detection and analysis. Summary of the invention

[0005] The present invention provides a method for processing defect images of autoclaved lightweight new wall materials, aiming to propose a processing model for defect images of autoclaved lightweight new wall materials, wherein a feature transformation module performs standardized processing on defect image features of autoclaved lightweight new wall materials to improve the expression ability of defect image features of autoclaved lightweight new wall materials, a feature adaptive processing module performs adaptive processing on different defect image features of autoclaved lightweight new wall materials to improve the matching degree and processing effect of defect image features of autoclaved lightweight new wall materials, a feature enhancement module comprehensively utilizes feature transformation and adaptive processing to improve the expression and processing accuracy of defect image features of autoclaved lightweight new wall materials, and a feature optimization module optimizes defect image features of autoclaved lightweight new wall materials to further improve the recognition degree and processing effect of features through channel dimension operations.

[0006] The present invention aims to propose an autoclaved lightweight new wall material defect image processing model and provide an autoclaved lightweight new wall material defect image processing method, comprising the following steps:

[0007] S1. Collect multiple autoclaved lightweight new wall material defect images, and pre-process the autoclaved lightweight new wall material defect images to form an autoclaved lightweight new wall material defect image dataset;

[0008] S2, construct a feature transformation module, including dimension transformation and matrix multiplication;

[0009] S3, constructing a feature adaptive processing module, including adaptive segmentation;

[0010] S4, constructing a feature enhancement module, including a feature transformation module and a feature adaptive processing module;

[0011] S5. Construct a feature optimization module, including channel dimension splitting and channel dimension splicing;

[0012] S6. Construct an autoclaved lightweight new wall material defect image processing model, including a feature enhancement module, a feature optimization module and a convolution layer;

[0013] S7. Training and processing of autoclaved lightweight new wall material defect image processing model. Use the autoclaved lightweight new wall material defect image dataset to train the autoclaved lightweight new wall material defect image processing model. After the training is completed, input a single autoclaved lightweight new wall material defect image to be processed into the autoclaved lightweight new wall material defect image processing model to obtain a single enhanced autoclaved lightweight new wall material defect image.

[0014] Preferably, in step S1, for the autoclaved lightweight new wall material defect image dataset, a plurality of clear autoclaved lightweight new wall material defect images are first collected, including photographing the defect area of ​​the autoclaved lightweight new wall material, and blurring the plurality of clear autoclaved lightweight new wall material defect images using Gaussian blur, defocus blur, motion blur and noise addition to obtain a plurality of blurred autoclaved lightweight new wall material defect images, and forming a plurality of training pairs with the corresponding plurality of clear autoclaved lightweight new wall material defect images, thereby forming an autoclaved lightweight new wall material defect image dataset.

[0015] Preferably, in step S2, for the feature transformation module, the defect feature of the autoclaved lightweight new wall material is input , , , and Representative defect characteristics of autoclaved lightweight new wall materials The height, width and channel of the autoclaved lightweight new wall material are characterized by defects Input to group normalization GN to get a feature , , for a feature Perform dimensional transformation and divide it into N feature blocks according to the feature block size value L. The height, width and channel of each feature block are L, L and 3. The combined dimension of the height and width of each feature block is , thus forming a secondary feature , ,in , using three weight matrices , and For secondary features Perform a linear transformation, , , ,in and Represents two different dimension values, and then gets the query ,key Sum , , , , , , , Represents matrix multiplication, query ,key Sum is the output of the feature transformation module.

[0016] Preferably, in step S2, for the feature transformation module, the dimensional transformation enables the autoclaved lightweight new wall material defect image processing model to effectively represent the autoclaved lightweight new wall material defect image features at different scales and capture more detail information. By segmenting and combining feature blocks, the diversity of features is increased, so that the autoclaved lightweight new wall material defect image processing model can better cope with different types of autoclaved lightweight new wall material defect images. The matrix multiplication process enhances the correlation between features, extracts more representative autoclaved lightweight new wall material defect image features, and improves the recognition and processing capabilities of the autoclaved lightweight new wall material defect image processing model.

[0017] Preferably, in step S3, for the feature adaptive processing module, the query output of the feature transformation module is obtained. ,key Sum , , , , N represents N feature blocks, Represents the dimension of each feature block, and queries and key The transposed matrix of Perform matrix multiplication, , get the features , , ,in Represents the feature block size value, represents the learnable offset value, , for the features Perform adaptive segmentation to obtain , and , , , ,in , and Respectively represent , and The channel, , , , , , and The specific calculation is as follows: First, initialize the scale parameter , and , , , , , and There are three learnable scale parameters, and then in training we get , and , , , and then get the features , , , Represents sparse maximization, and then obtains the features , , , represents the scaled exponential linear unit, Representatives will The output is squared and then the feature is obtained , , , Represents splicing in the channel dimension, represents a linear transformation, Representative value , then the feature Perform dimension transformation to obtain features , ,feature is the output of the feature adaptation processing module.

[0018] Preferably, in step S3, for the feature adaptive processing module, adaptive segmentation enables the autoclaved lightweight new wall material defect image processing model to more flexibly process features of different types and scales. By using learnable parameters, the autoclaved lightweight new wall material defect image processing model can be adaptively adjusted according to specific input features, thereby improving the effect and accuracy of feature processing. The application of sparse operations and nonlinear activation functions makes the feature representation more sparse and effective, reduces redundant information, and improves the processing efficiency of the autoclaved lightweight new wall material defect image processing model.

[0019] Preferably, in step S4, the feature enhancement module includes a feature transformation module and a feature adaptive processing module, and inputs the defect feature of the autoclaved lightweight new wall material. To the feature transformation module, , , and Representative defect characteristics of autoclaved lightweight new wall materials The height, width and channel of the query are obtained ,key Sum , then query ,key Sum Input to the feature adaptive processing module to obtain the feature , , then the feature and Features Add element by element to get features , ,feature is the output of the feature enhancement module.

[0020] Preferably, in step S4, for the feature enhancement module, by innovatively combining the feature transformation module and the feature adaptive processing module, and adding the original features and the processed features element by element in the final step, not only the expressive power and representation effect of the features are improved, but also the process and efficiency of feature processing are optimized.

[0021] Preferably, in step S5, for the feature optimization module, input feature , , , and Representative features The height, width and channel of , ,feature The two dimensions are and , and The value of is different. , represents group normalization, represents masked convolution, represents the first linear layer, and then the features Split the channel dimension to obtain features and Features , , , the feature Perform dimension transformation to obtain features , , the feature Input to the dilated convolution to obtain features , , then the feature Perform dimension transformation to obtain features , , the feature and Features According to weight and Perform channel dimension splicing to obtain features , , , weight and The calculation method is , , and are two learnable parameters, which transform the features Input to the second linear layer to get the features , , the feature and Features Add element by element to get features , ,feature is the output of the feature optimization module.

[0022] Preferably, in step S5, for the feature optimization module, the features are split and spliced ​​in the channel dimension to improve the diversity and expression ability of the features, and the void convolution is used to enhance the spatial information capture ability of the features, so that the autoclaved lightweight new wall material defect image processing model can better process large-scale image information, and the optimized features are added element by element to the original features to enhance the representation ability of the features and the robustness of the autoclaved lightweight new wall material defect image processing model.

[0023] Preferably, in step S6, for the autoclaved lightweight new wall material defect image processing model, the number of Multiple series feature enhancement modules and a convolutional layer are connected in series to form a type of enhancement block. The number of serial feature enhancement modules and Multiple series feature optimization modules are connected in series to form a type II enhancement block, and then a single fuzzy autoclaved lightweight new wall material defect image is input. to the defect image processing model of autoclaved lightweight new wall material, , , and 3 represent a single fuzzy autoclaved lightweight new wall material defect image The height, width and channel of the single blurred autoclaved lightweight new wall material defect image Input to the first convolutional layer to get the features , , , and Representative features The height, width and channel of Enter the quantity as Multiple series-connected type-one enhancement blocks are used to obtain the characteristic , , the feature Enter the quantity as Multiple series-connected type II enhancement blocks are used to obtain the characteristic , , the feature Input to the second convolution layer to get the intermediate image , , the intermediate image and a single blurred autoclaved lightweight new wall material defect image Add element by element to get a single enhanced autoclaved lightweight new wall material defect image , , single enhanced autoclaved lightweight new wall material defect image It is the output of the defect image processing model of autoclaved lightweight new wall material.

[0024] Compared with the prior art, the present invention has the following technical effects: the technical solution provided by the present invention proposes a defect image processing model for autoclaved lightweight new wall materials, wherein the feature transformation module performs standardized processing on image features to improve the expressiveness of image features, the feature adaptive processing module performs adaptive processing on different image features to improve the matching degree and processing effect of image features, the feature enhancement module comprehensively utilizes feature transformation and adaptive processing to improve the expressiveness and processing accuracy of image features, and the feature optimization module optimizes image features and further improves the feature recognition and processing effect through channel dimension operations. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a defect image processing flow chart of the autoclaved lightweight new wall material provided by the present invention.

[0026] Figure 2 It is a structural diagram of the defect image processing model of the autoclaved lightweight new wall material provided by the present invention.

[0027] Figure 3 It is a single blurred autoclaved lightweight new wall material defect image provided by the present invention.

[0028] Figure 4 It is a defect image of a single reinforced autoclaved lightweight new wall material provided by the present invention. DETAILED DESCRIPTION

[0029] The present invention aims to propose a method for processing defect images of autoclaved lightweight new wall materials, and proposes a defect image processing model for autoclaved lightweight new wall materials, wherein a feature transformation module performs standardized processing on image features to improve the expressiveness of image features, a feature adaptive processing module performs adaptive processing on different image features to improve the matching degree and processing effect of image features, a feature enhancement module comprehensively utilizes feature transformation and adaptive processing to improve the expressiveness and processing accuracy of image features, and a feature optimization module optimizes image features to further improve the recognition and processing effect of features through operations in channel dimensions.

[0030] See also Figure 1 As shown, a defect image processing method of an autoclaved lightweight new wall material in an embodiment of the present application includes specific steps S1 to S7.

[0031] S1. Collect multiple autoclaved lightweight new wall material defect images, and pre-process the autoclaved lightweight new wall material defect images to form an autoclaved lightweight new wall material defect image dataset.

[0032] Furthermore, for the autoclaved lightweight new wall material defect image dataset, first, multiple clear autoclaved lightweight new wall material defect images are collected, including taking images of defect areas of the autoclaved lightweight new wall material, and using Gaussian blur, defocus blur, motion blur and noise addition to blur the multiple clear autoclaved lightweight new wall material defect images to obtain multiple blurred autoclaved lightweight new wall material defect images, and forming multiple training pairs with the corresponding multiple clear autoclaved lightweight new wall material defect images, thereby forming an autoclaved lightweight new wall material defect image dataset.

[0033] S2. Construct a feature transformation module, including dimension transformation and matrix multiplication.

[0034] Furthermore, for the feature transformation module, the defect features of the autoclaved lightweight new wall material are input , , , and Representative defect characteristics of autoclaved lightweight new wall materials The height, width and channel of the autoclaved lightweight new wall material are characterized by defects Input to group normalization GN to get a feature , , for a feature Perform dimensional transformation and divide it into N feature blocks according to the feature block size value L. The height, width and channel of each feature block are L, L and 3. The combined dimension of the height and width of each feature block is , thus forming a secondary feature , ,in , using three weight matrices , and For secondary features Perform a linear transformation, , , ,in and Represents two different dimension values, and then gets the query ,key Sum , , , , , , , Represents matrix multiplication, query ,key Sum is the output of the feature transformation module.

[0035] S3. Construct a feature adaptive processing module, including adaptive segmentation.

[0036] Further, for the feature adaptive processing module, the query output of the feature transformation module is obtained ,key Sum , , , , N represents N feature blocks, Represents the dimension of each feature block, and queries and key The transposed matrix of Perform matrix multiplication, , get the features , , ,in Represents the feature block size value, represents the learnable offset value, , for the features Perform adaptive segmentation to obtain , and , , , ,in , and Respectively represent , and The channel, , , , , , and The specific calculation is as follows: First, initialize the scale parameter , and , , , , , and There are three learnable scale parameters, and then in training we get , and , , , and then get the features , , , Represents sparse maximization, and then obtains the features , , , represents the scaled exponential linear unit, Representatives will The output is squared and then the feature is obtained , , , Represents splicing in the channel dimension, represents a linear transformation, Representative value , then the feature Perform dimension transformation to obtain features , ,feature is the output of the feature adaptation processing module.

[0037] S4. Construct a feature enhancement module, including a feature transformation module and a feature adaptive processing module.

[0038] Furthermore, for the feature enhancement module, including the feature transformation module and the feature adaptive processing module, the defect features of the autoclaved lightweight new wall material are input To the feature transformation module, , , and Representative defect characteristics of autoclaved lightweight new wall materials The height, width and channel of the query are obtained ,key Sum , then query ,key Sum Input to the feature adaptive processing module to obtain the feature , , then the feature and Features Add element by element to get features , ,feature is the output of the feature enhancement module.

[0039] S5. Construct a feature optimization module, including channel dimension splitting and channel dimension splicing.

[0040] Furthermore, for the feature optimization module, the input feature , , , and Representative features The height, width and channel of , ,feature The two dimensions are and , and The value of is different. , represents group normalization, represents masked convolution, represents the first linear layer, and then the features Split the channel dimension to obtain features and Features , , , the feature Perform dimension transformation to obtain features , , the feature Input to the dilated convolution to obtain features , , then the feature Perform dimension transformation to obtain features , , the feature and Features According to weight and Perform channel dimension splicing to obtain features , , , weight and The calculation method is , , and are two learnable parameters, which transform the features Input to the second linear layer to get the features , , the feature and Features Add element by element to get features , ,feature is the output of the feature optimization module.

[0041] S6. Construct a defect image processing model for autoclaved lightweight new wall materials, including a feature enhancement module, a feature optimization module and a convolutional layer.

[0042] Furthermore, for the defect image processing model of autoclaved lightweight new wall material, the number of Multiple series feature enhancement modules and a convolutional layer are connected in series to form a type of enhancement block. The number of serial feature enhancement modules and Multiple series feature optimization modules are connected in series to form a type II enhancement block, and then a single fuzzy autoclaved lightweight new wall material defect image is input. to the defect image processing model of autoclaved lightweight new wall material, , , and 3 represent a single fuzzy autoclaved lightweight new wall material defect image The height, width and channel of the single blurred autoclaved lightweight new wall material defect image Input to the first convolutional layer to get the features , , , and Representative features The height, width and channel of Enter the quantity as Multiple series-connected type-one enhancement blocks are used to obtain the characteristic , , the feature Enter the quantity as Multiple series-connected type II enhancement blocks are used to obtain the characteristic , , the feature Input to the second convolution layer to get the intermediate image , , the intermediate image and a single blurred autoclaved lightweight new wall material defect image Add element by element to get a single enhanced autoclaved lightweight new wall material defect image , , single enhanced autoclaved lightweight new wall material defect image It is the output of the defect image processing model of autoclaved lightweight new wall material.

[0043] S7. Training and processing of autoclaved lightweight new wall material defect image processing model. Use the autoclaved lightweight new wall material defect image dataset to train the autoclaved lightweight new wall material defect image processing model. After the training is completed, input a single autoclaved lightweight new wall material defect image to be processed into the autoclaved lightweight new wall material defect image processing model to obtain a single enhanced autoclaved lightweight new wall material defect image.

[0044] Furthermore, the autoclaved lightweight new wall material defect image processing model is written in Python language, using the Pytorch framework. The code runs on the Linux system, the version is Ubuntu 20.04 LTS, using Nvidia RTX4080 graphics card, using CUDA and cuDNN, and the default settings of the AdamW optimizer are used to train the autoclaved lightweight new wall material defect image processing model. The number of training rounds is 400, and the learning rate is initially set to 0.0001, and then gradually reduced to 0.000001 using the cosine decay strategy.

[0045] Furthermore, a single blurred autoclaved lightweight new wall material defect image such as Figure 3 As shown in FIG. 1 , a single fuzzy autoclaved lightweight new wall material defect image is input into the autoclaved lightweight new wall material defect image processing model, and the following is obtained: Figure 4 The single enhanced autoclaved lightweight new wall material defect image shown in the figure is compared with the single blurred autoclaved lightweight new wall material defect image. It can be seen that after being processed by the autoclaved lightweight new wall material defect image processing model, the clarity and details of the single enhanced autoclaved lightweight new wall material defect image are greatly improved.

[0046] The above are only preferred embodiments of the present invention. It should be pointed out that a person skilled in the art can make several modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A defect image processing method for autoclaved lightweight new wall material, characterized in that: The following steps are involved: S1. Collect multiple autoclaved lightweight new wall material defect images, and pre-process the autoclaved lightweight new wall material defect images to form an autoclaved lightweight new wall material defect image dataset; S2. Construct a feature transformation module, including dimension transformation and matrix multiplication; the feature transformation module inputs the defect features of the autoclaved lightweight new wall material , , , and Representative defect characteristics of autoclaved lightweight new wall materials The height, width and channel of the autoclaved lightweight new wall material are characterized by defects Input to group normalization GN to get a feature , , for a feature Perform dimensional transformation and divide it into N feature blocks according to the feature block size value L. The height, width and channel of each feature block are L, L and 3. The combined dimension of the height and width of each feature block is , thus forming a secondary feature , ,in , using three weight matrices , and For secondary features Perform a linear transformation, , , ,in and Represents two different dimension values, and then gets the query ,key Sum , , , , , , , Represents matrix multiplication, query ,key Sum is the output of the feature transformation module; S3, constructing a feature adaptive processing module, including adaptive segmentation; S4, constructing a feature enhancement module, including a feature transformation module and a feature adaptive processing module; S5. Construct a feature optimization module, including channel dimension splitting and channel dimension splicing; S6. Construct an autoclaved lightweight new wall material defect image processing model, including a feature enhancement module, a feature optimization module and a convolution layer; S7. Training and processing of autoclaved lightweight new wall material defect image processing model. Use the autoclaved lightweight new wall material defect image dataset to train the autoclaved lightweight new wall material defect image processing model. After the training is completed, input a single autoclaved lightweight new wall material defect image to be processed into the autoclaved lightweight new wall material defect image processing model to obtain a single enhanced autoclaved lightweight new wall material defect image.

2. The defect image processing method of the autoclaved lightweight new wall material according to claim 1 is characterized in that: In the step S1, for the autoclaved lightweight new wall material defect image dataset, firstly, a plurality of clear autoclaved lightweight new wall material defect images are collected, images of defect areas of the autoclaved lightweight new wall material are photographed, and the plurality of clear autoclaved lightweight new wall material defect images are blurred using Gaussian blur, defocus blur, motion blur and noise addition to obtain a plurality of blurred autoclaved lightweight new wall material defect images, and a plurality of training pairs are formed corresponding to the plurality of clear autoclaved lightweight new wall material defect images, thereby forming an autoclaved lightweight new wall material defect image dataset.

3. The defect image processing method of the autoclaved lightweight new wall material according to claim 1 is characterized in that: In the step S3, for the feature adaptive processing module, the query output of the feature transformation module is obtained. ,key Sum , , , , N represents N feature blocks, Represents the dimension of each feature block, and queries and key The transposed matrix of Perform matrix multiplication, , get the features , , ,in Represents the feature block size value, represents the learnable offset value, , for the features Perform adaptive segmentation to obtain , and , , , ,in , and Respectively represent , and The channel, , , , , , and The specific calculation is as follows: First, initialize the scale parameter , and , , , , , and There are three learnable scale parameters, and then in training we get , and , , , and then get the features , , , Represents sparse maximization, and then obtains the features , , , represents the scaled exponential linear unit, Representatives will The output is squared and then the feature is obtained , , , Represents splicing in the channel dimension, represents a linear transformation, Representative value , then the feature Perform dimension transformation to obtain features , ,feature is the output of the feature adaptation processing module.

4. The defect image processing method of the autoclaved lightweight new wall material according to claim 1 is characterized in that: In the step S4, the feature enhancement module includes a feature transformation module and a feature adaptive processing module, and inputs the defect feature of the autoclaved lightweight new wall material. To the feature transformation module, , , and Representative defect characteristics of autoclaved lightweight new wall materials The height, width and channel of the query are obtained ,key Sum , then query ,key Sum Input to the feature adaptive processing module to obtain the feature , , then the feature and Features Add element by element to get features , ,feature is the output of the feature enhancement module.

5. The defect image processing method of the autoclaved lightweight new wall material according to claim 1 is characterized in that: In the step S5, for the feature optimization module, input feature , , , and Representative features The height, width and channel of , ,feature The two dimensions are and , and The value of is different. , represents group normalization, represents masked convolution, represents the first linear layer, and then the features Split the channel dimension to obtain features and Features , , , the feature Perform dimension transformation to obtain features , , the feature Input to the dilated convolution to obtain features , , then the feature Perform dimension transformation to obtain features , , the feature and Features According to weight and Perform channel dimension splicing to obtain features , , , weight and The calculation method is , , and are two learnable parameters, which transform the features Input to the second linear layer to get the features , , the feature and Features Add element by element to get features , ,feature is the output of the feature optimization module.

6. The defect image processing method of the autoclaved lightweight new wall material according to claim 1 is characterized in that: In the step S6, for the autoclaved lightweight new wall material defect image processing model, the number is Multiple series feature enhancement modules and a convolutional layer are connected in series to form a type of enhancement block. The number of serial feature enhancement modules is Multiple series feature optimization modules are connected in series to form a type II enhancement block, and then a single fuzzy autoclaved lightweight new wall material defect image is input. to the defect image processing model of autoclaved lightweight new wall material, , , and 3 represent a single fuzzy autoclaved lightweight new wall material defect image The height, width and channel of the single blurred autoclaved lightweight new wall material defect image Input to the first convolutional layer to get the features , , , and Representative features The height, width and channel of Enter the quantity as Multiple series-connected type-one enhancement blocks are used to obtain the characteristic , , the feature Enter the quantity as Multiple series-connected type II enhancement blocks are used to obtain the characteristic , , the feature Input to the second convolution layer to get the intermediate image , , the intermediate image and a single blurred autoclaved lightweight new wall material defect image Add element by element to get a single enhanced autoclaved lightweight new wall material defect image , , single enhanced autoclaved lightweight new wall material defect image It is the output of the defect image processing model of autoclaved lightweight new wall material.

Images