A method for identifying defects in ALC green wall panels

Abstract

The present invention proposes an ALC green environmentally friendly wallboard defect recognition method, which relates to the field of image recognition. The present invention proposes an ALC green environmentally friendly wallboard defect recognition process, including ALC green environmentally friendly wallboard defect data set production, multi-head self-attention construction, pixel focus block construction, feature transfer feedforward network construction, efficient feature extraction block construction, efficient feature enhancement module construction, ALC green environmentally friendly wallboard defect recognition model construction and training ALC green environmentally friendly wallboard defect recognition model and detection; at the same time, the proposed multi-head self-attention captures long-distance information through the dependency relationship between global features, the pixel focus block focuses on local feature extraction, enhances the model's ability to capture position and details, and the feature transfer feedforward network strengthens important features through adaptive channel weighting to improve feature expression capabilities.

Description

Technical Field

[0001] The invention belongs to the field of image recognition, and in particular relates to an ALC green environmentally friendly wallboard defect recognition method. Background Art

[0002] ALC green and environmentally friendly wall panels are widely used due to their light weight, thermal insulation and sound insulation, but cracks, bubbles, discoloration, peeling, stains and dents may occur during the production process, affecting their performance and appearance; traditional defect detection methods such as visual inspection rely on manual operation and are easily affected by subjective factors, resulting in missed detection or misjudgment, and the inefficient inspection speed is difficult to meet production needs; ALC green and environmentally friendly wall panel defect identification method improves detection efficiency and accuracy through automation and intelligent means, significantly shortens detection time, reduces dependence on manual labor, thereby reducing labor costs, and promptly discovers product quality problems, improving product reliability and consistency.

[0003] Traditional convolutional neural networks (CNNs) are limited to local receptive fields and have difficulty capturing global contextual information. Although Transformers perform well in global understanding, they have low computational efficiency and may ignore details. Combining the advantages of CNNs and Transformers can simultaneously achieve accurate capture of local features and effective understanding of the global context, thereby improving the computational efficiency and detail retention capabilities of the model. This combination reduces computational redundancy and improves the accuracy of defect recognition in ALC green wall panels. Summary of the invention

[0004] The present invention provides an ALC green environmentally friendly wallboard defect recognition method, aiming to propose multi-head self-attention, pixel focusing block and feature transfer feedforward network; multi-head self-attention enhances the ability to capture diversified features and information relationships through multiple independent attention heads, so that the model can take into account both global context and local details; the pixel focusing block uses convolution operation to extract local features, and enhances spatial perception through position embedding, so that the model can process long sequence data; the feature transfer feedforward network effectively fuses local and global information, and improves the performance of the model in complex tasks; the modules work together to ensure that the defect recognition model achieves excellent detection performance while maintaining high computational efficiency.

[0005] The present invention aims to propose a multi-head self-attention, pixel focusing block, feature transfer feedforward network, efficient feature extraction block and efficient feature enhancement module, and provide an ALC green environmentally friendly wallboard defect recognition method, comprising the following steps:

[0006] S1. ALC green environmentally friendly wallboard defect dataset is prepared. ALC green environmentally friendly wallboard defect images are collected, including various defect types. Defects are annotated for each ALC green environmentally friendly wallboard defect image to form an ALC green environmentally friendly wallboard defect dataset.

[0007] S2. Construct a multi-head self-attention network, split the feature map into N feature subspaces and perform self-attention operations on them respectively. Then, connect the feature maps output by the self-attention operation and perform depth-wise separable convolution. Finally, perform residual connection and output the feature map.

[0008] S3, construct a pixel focusing block, capture the local feature information of the feature map through flattening, position embedding, normalization, activation, multi-channel 1D convolution, and perform residual connection, and finally output the feature map;

[0009] S4. Construct a feature transfer feed-forward network, including global average pooling, two fully connected layers, broadcast context information, and perform residual connection and output;

[0010] S5. Construct an efficient feature extraction block, including N efficient feature extraction modules, which are composed of multi-head self-attention, pixel focusing block and feature transfer feedforward network;

[0011] S6. Build an efficient feature enhancement module, including 1×1 convolution, 5×5 convolution, and 5 efficient feature extraction blocks, and output the enhanced feature map;

[0012] S7, constructing an ALC green environmentally friendly wallboard defect recognition model, including input, efficient feature enhancement module, ALC green environmentally friendly wallboard defect detection head and output;

[0013] S8. Train the ALC green environmentally friendly wall panel defect recognition model and detect it. Use the ALC green environmentally friendly wall panel defect data set to train the ALC green environmentally friendly wall panel defect recognition model. After the training is completed, input the new ALC green environmentally friendly wall panel photo to be detected into the ALC green environmentally friendly wall panel defect recognition model to obtain the detection result. The detection result includes the defect location and type in the ALC green environmentally friendly wall panel picture.

[0014] Preferably, in step S1, the various defect types include cracks, bubbles, discoloration, peeling, stains and depressions.

[0015] Preferably, in step S2, for multi-head self-attention, the input feature , , , and Respectively represent The height, width and channel of Split into N features, we get , , Represents the features of the ith split, and N heads calculate attention independently. , Represent the linear transformation weight matrix of the i-th head, Represent the query, key and value of the i-th header respectively. and Perform adaptive pooling, , , is the fixed size after pooling, is the number of channels of the i-th split feature, and the attention weight is calculated for each head. , , is the attention weight matrix of the ith head, is a scaling factor equal to the dimension of the key vector, The key to be applied to the pool , , , is the weighted output of the i-th head, and performs residual connection and convolution operations to further capture local feature information. , , represents a 1×1 convolution, which concatenates the outputs of all heads, focuses on the features of different subspaces, and restores the original channel dimension. , , the concatenated output features are processed through depth-wise separable convolution to reduce the number of parameters and improve computational efficiency, , Finally, the input features are added to the processed features through residual connection, retaining the original information of the input features to obtain the final output. , .

[0016] Preferably, in step S2, for multi-head self-attention, by calculating the dependencies between features globally and capturing the long-distance relationships of features from multiple perspectives, different attention heads can focus on different parts of the feature map, enriching the context-awareness of the model and enabling the global information to be effectively integrated.

[0017] Preferably, in step S3, for the pixel focus block, input , , flatten the input features, convert spatial information into sequence information, and use the sequence model's sensitivity to position to capture information. , , for ,Add position embedding to the flattened features to enhance position awareness and avoid losing spatial structure of serialized features after flattening, , , Represents a length The sequence position embedding is used to adjust the feature distribution through normalization to improve the stability of the model. , , Represents the normalization operation, which is activated by the LeakyReLU function to introduce nonlinearity in the serialized features to better express complex features. , , It represents the activation function, which extracts features in the sequence dimension on multiple feature channels through multi-channel 1D convolution, captures the correlation between features at adjacent positions, and performs residual connection to avoid gradient disappearance. , , represents multi-channel 1D convolution, Represents a connection operation, Stands for element-wise addition.

[0018] Preferably, in step S3, for the pixel focusing block, the pixel focusing block focuses on local features, enhances position perception through position embedding, captures neighborhood features using multi-channel 1D convolution, and retains input information using residual connection.

[0019] Preferably, in step S4, for the feature transfer feedforward network, the input feature , , the features are pooled globally to retain the overall features of each channel, , , It stands for global average pooling, which generates adaptive channel weights through two fully connected layers, adds nonlinear expression, and adaptively adjusts the weights of each channel. , , , is the weight matrix, is the scaling factor, represents the bias term, Represents the activation function, broadcasting the generated weights to the spatial dimensions , multiply the broadcast weights by the input feature map channel by channel, so that the weight of each channel affects the entire feature map, and perform residual connections to avoid gradient disappearance to obtain the output , , , represents channel-by-channel multiplication, Stands for channel-by-channel, element-by-element addition.

[0020] Preferably, in step S4, for the feature transfer feedforward network, important features are adaptively emphasized through a channel weighting mechanism, weights are generated using global average pooling and fully connected layers to enhance significant features, feature expression is optimized at the channel level, and features are robustly transferred through residual connections to enhance the nonlinear expression capability of the model.

[0021] Preferably, in step S5, for the efficient feature extraction block, the input feature , , stacking N efficient feature extraction modules, integrating long-distance dependent global features and fine-grained local features to improve the performance and efficiency of the model, , , , Representative An efficient feature extraction module, , represents multi-head self-attention, represents the pixel focus block, Represents feature transfer feed-forward network.

[0022] Preferably, in step S5, for the efficient feature extraction block, by stacking multiple efficient feature extraction modules, the long-distance dependent global features and fine-grained local features are effectively fused, so that the model has global perception capabilities while retaining detail information; wherein each module utilizes the features extracted by multi-head self-attention and pixel focusing blocks, and further optimizes and enhances the feature expression through the feature transfer feedforward network, thereby improving the performance and computational efficiency of the model.

[0023] Preferably, in step S6, for the efficient feature enhancement module, the input feature , , obtained by capturing local information through the 5×5 convolution enhancement model and the first efficient feature extraction block , , by downsampling to reduce the spatial size and computational complexity, and further extracting deep features through the second efficient feature extraction block , , further reduce the spatial resolution by downsampling, extract deeper features, and further extract deep features through the third efficient feature extraction block to obtain , , the feature information between different resolution layers is retained by upsampling, the residual connection retains the middle-level feature expression, and then the features are fused through 1×1 convolution, and the deep features are further extracted through the fourth efficient feature extraction block. , , restored to the original resolution by upsampling, the residual connection retains the initial feature information, the 1×1 convolution fusion feature and the fifth efficient feature extraction block extract the final enhanced features , Finally, a 5×5 convolution is performed on the final feature map to further smooth and integrate the output , , , Represents a 5×5 convolution operation.

[0024] Preferably, in step S6, for the efficient feature enhancement module, global and local information is gradually extracted through multi-level feature extraction and fusion: the initial input is refined through 5×5 convolution and a series of efficient feature extraction blocks, and features of different resolutions are gradually fused through multi-layer downsampling and upsampling, and residual connections are combined to retain key information. Finally, the features are further integrated through 1×1 convolution and 5×5 convolution to generate an enhanced output with both details and semantic information.

[0025] Preferably, in step S7, for the ALC green wallboard defect recognition model, the input image , image features are obtained through the backbone network , input the efficient feature enhancement module to obtain the enhanced output features , and then use the detection head to detect and obtain the ALC green environmentally friendly wall panel defect detection results, which include the location and type of the ALC green environmentally friendly wall panel defects.

[0026] Compared with the prior art, the present invention has the following technical effects:

[0027] The technical solution provided by the present invention proposes multi-head self-attention, pixel focusing block, and feature transfer feedforward network; multi-head self-attention calculates the dependencies between features in a global scope and captures the long-distance relationship of features from multiple perspectives. Different attention heads can focus on different parts of the feature map, enriching the context perception ability of the model and effectively integrating global information; the pixel focusing block focuses on local features, enhances position perception ability through position embedding, captures neighborhood features using 1D convolution, and retains input information using residual connection. This design enables the model to capture local details finely, especially performing well on fine-grained information such as edges and textures; the feature transfer feedforward network adaptively emphasizes important features through a channel weighting mechanism, uses global average pooling and fully connected layers to generate weights to enhance significant features, optimizes feature expression at the channel level, and robustly transfers features through residual connections to improve the nonlinear expression ability of the model; the combined application of multiple modules ensures a good balance between computational efficiency and detection performance in the ALC green environmentally friendly wallboard defect model. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 This is a defect identification flow chart of ALC green environmentally friendly wallboard provided by the present invention.

[0029] Figure 2 This is a multi-head self-attention structure diagram provided by the present invention.

[0030] Figure 3 This is a structural diagram of a pixel focusing block provided by the present invention.

[0031] Figure 4 This is a structural diagram of the feature transfer feedforward network provided by the present invention.

[0032] Figure 5 This is a block structure diagram of efficient feature extraction provided by the present invention.

[0033] Figure 6 This is a structural diagram of the ALC green environmentally friendly wallboard defect recognition model provided by the present invention.

[0034] Figure 7 This is the detection result of the ALC green environmentally friendly wallboard defect recognition model provided by the present invention. DETAILED DESCRIPTION

[0035] The present invention aims to propose a method for identifying defects in ALC green environmentally friendly wall panels, and proposes multi-head self-attention, pixel focusing blocks and feature transfer feedforward networks; multi-head self-attention enhances the ability to capture diversified features and information relationships through multiple independent attention heads, so that the model can take into account both global context and local details; the pixel focusing block uses convolution operations to extract local features, and enhances spatial perception through position embedding, so that the model can process long sequence data; the feature transfer feedforward network effectively fuses local and global information and improves the performance of the model in complex tasks; the modules are used in a coordinated manner to ensure that the ALC green environmentally friendly wall panel defect recognition model achieves excellent detection performance while maintaining high computational efficiency.

[0036] See also Figure 1 As shown, a method for identifying defects of ALC green environmentally friendly wallboards in an embodiment of the present application:

[0037] S1. ALC green environmentally friendly wallboard defect dataset was prepared. A total of 1,000 images of ALC green environmentally friendly wallboard defects were collected, including various defect types: cracks, bubbles, discoloration, peeling, stains and dents. Each ALC green environmentally friendly wallboard defect image was annotated to form the ALC green environmentally friendly wallboard defect dataset.

[0038] Furthermore, in step S1, various defect types are detected, including cracks, bubbles, discoloration, peeling, stains and depression defects.

[0039] S2. Construct multi-head self-attention, split the feature map into 4 feature subspaces and perform self-attention operations on them respectively. Then connect the feature maps output by the self-attention operation and perform depth-wise separable convolution. Finally, perform residual connection and output the feature map.

[0040] Furthermore, in step S2, for multi-head self-attention, its structure is as follows Figure 2 As shown, the input features , , H, W and C represent The height, width and channel of Split into 4 features, we get , , Represents the features of the i-th split, and calculates the attention of the four heads independently. , Represent the linear transformation weight matrix of the i-th head, Represent the query, key and value of the i-th header respectively. and Perform adaptive pooling. , , is the fixed size after pooling, is the number of channels of the i-th split feature, and the attention weight is calculated for each head. , , is the attention weight matrix of the ith head, is a scaling factor equal to the dimension of the key vector, The key to be applied to the pool , , , is the weighted output of the i-th head, and performs residual connection and convolution operations to further capture local feature information. , , represents a 1×1 convolution, which concatenates the outputs of all heads, focuses on the features of different subspaces, and restores the original channel dimension. , , the concatenated output features are processed through depth-wise separable convolution to reduce the number of parameters and improve computational efficiency, , Finally, the input features are added to the processed features through residual connection, retaining the original information of the input features to obtain the final output. , .

[0041] S3. Construct a pixel focusing block to capture the local feature information of the feature map through flattening, position embedding, normalization, activation, multi-channel 1D convolution, and perform residual connection, and finally output the feature map.

[0042] Further, in step S3, for the pixel focus block, its structure is as follows Figure 3 As shown, input , , flatten the input features, convert spatial information into sequence information, and use the sequence model's sensitivity to position to capture information. , , for ,Add position embedding to the flattened features to enhance position awareness and avoid losing spatial structure of serialized features after flattening, , , Represents a length The sequence position embedding is used to adjust the feature distribution through normalization to improve the stability of the model. , , Represents the normalization operation, which is activated by the LeakyReLU function to introduce nonlinearity in the serialized features to better express complex features. , , It represents the activation function, which extracts features in the sequence dimension on multiple feature channels through multi-channel 1D convolution, captures the correlation between features at adjacent positions, and performs residual connection to avoid gradient disappearance. , , represents multi-channel 1D convolution, Represents a connection operation, Stands for element-wise addition.

[0043] S4. Construct a feature transfer feedforward network, including global average pooling, two fully connected layers, broadcast context information, and perform residual connection and output.

[0044] Further, in step S4, for the feature transfer feedforward network, its structure is as follows Figure 4 As shown, the input features , , the features are pooled globally to retain the overall features of each channel, , , It stands for global average pooling, which generates adaptive channel weights through two fully connected layers, adds nonlinear expression, and adaptively adjusts the weights of each channel. , , , is the weight matrix, is the scaling factor, represents the bias term, Represents the activation function, broadcasting the generated weights to the spatial dimensions , multiply the broadcast weights by the input feature map channel by channel, so that the weight of each channel affects the entire feature map, and perform residual connections to avoid gradient disappearance to obtain the output , , , represents channel-by-channel multiplication, Stands for channel-by-channel, element-by-element addition.

[0045] S5. Build an efficient feature extraction block, including 5 efficient feature extraction modules. The efficient feature extraction module is composed of multi-head self-attention, pixel focusing block and feature transfer feedforward network.

[0046] Further, in step S5, for the efficient feature extraction block, its structure is as follows Figure 5 As shown, the input features , , stacking 5 efficient feature extraction modules, integrating long-distance dependent global features and fine-grained local features to improve the performance and efficiency of the model, , , , Representative An efficient feature extraction module, , represents multi-head self-attention, represents the pixel focus block, Represents feature transfer feed-forward network.

[0047] S6. Build an efficient feature enhancement module, including 1×1 convolution, 5×5 convolution, and 5 efficient feature extraction blocks, and output the enhanced feature map.

[0048] S7. Construct an ALC green environmentally friendly wall panel defect recognition model, including input, efficient feature enhancement module, ALC green environmentally friendly wall panel defect detection head and output.

[0049] Further, in step S7, for the ALC green environmental protection wallboard defect recognition model, its structure is as follows Figure 6 As shown, the input image , , image features are obtained through the backbone network , , input the efficient feature enhancement module, and obtain the local information captured by the 5×5 convolution enhancement model and the first efficient feature extraction block , , by downsampling to reduce the spatial size and computational complexity, and further extracting deep features through the second efficient feature extraction block , , further reduce the spatial resolution by downsampling, extract deeper features, and further extract deep features through the third efficient feature extraction block to obtain , , the feature information between different resolution layers is retained by upsampling, the residual connection retains the middle-level feature expression, and then the features are fused through 1×1 convolution, and the deep features are further extracted through the fourth efficient feature extraction block. , , restored to the original resolution by upsampling, the residual connection retains the initial feature information, the 1×1 convolution fusion feature and the fifth efficient feature extraction block extract the final enhanced features , Finally, a 5×5 convolution is performed on the final feature map to further smooth and integrate the output features. , , , Represents a 5×5 convolution operation to obtain enhanced output features , trained using the ALC green wallboard defect dataset.

[0050] S8. Train the ALC green environmentally friendly wall panel defect recognition model and detect it. Use the ALC green environmentally friendly wall panel defect data set to train the ALC green environmentally friendly wall panel defect recognition model. After the training is completed, input the new ALC green environmentally friendly wall panel photo to be detected into the ALC green environmentally friendly wall panel defect recognition model to obtain the detection result. The detection result includes the defect location and type in the ALC green environmentally friendly wall panel picture.

[0051] Further, in step S8, the ALC green environmentally friendly wall panel defect recognition model is trained. The defect recognition model is trained using the ALC green environmentally friendly wall panel defect dataset. The code is written using the Pycharm application and Python language. The Pytorch framework is used. The model inputs 640×640×3 ALC green environmentally friendly wall panel defect images for training. The model is trained from scratch for 250 epochs and optimized using the AdamW optimizer. The learning rate is set to 0.0003 and the batchsize is set to 64.

[0052] Furthermore, in step S8, the video is shot in real time, frames are extracted and input into the ALC green environmentally friendly wallboard defect recognition model. If there are defects on the surface, the detection model will identify the defects and mark them with a rectangular frame.

[0053] Further, in step S8, the ALC green environmental protection wallboard defect detection results are as follows: Figure 7 As shown, Figure 7 The defect detection box and detection confidence level are shown as 0.92. Figure 7 It can be seen that the defect detection model can successfully detect the wall panel defects.

[0054] 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 method for identifying defects in ALC green environmentally friendly wallboards, characterized in that: The following steps are involved: S1. ALC green environmentally friendly wallboard defect data set is produced. ALC green environmentally friendly wallboard defect pictures are collected, including various defect types. Each ALC green environmentally friendly wallboard defect picture is defect-labeled to form an ALC green environmentally friendly wallboard defect data set; S2. Multi-head self-attention is constructed to split the feature map into N feature subspaces for self-attention operations respectively, and then the feature maps output by the self-attention operation are connected and deep separable convolution is performed, and finally residual connection is performed and the feature map is output; In step S2, construct a multi-head self-attention and input feature X Attin , X Attin ∈R H×W×C , H, W and C represent X Attin The height, width and channel of X Attin Split into N features, we get Represents the features of the ith split, and N heads calculate attention independently. Represents the linear transformation weight matrix of the i-th head, Q i , K i , V i Represent the query, key and value of the i-th head respectively, for K i and V i Perform adaptive pooling, K i ′=AdaptivePool(K i ), V i ′=AdaptivePool(V i ), p is the fixed size after pooling, is the number of channels of the i-th split feature, and the attention weight is calculated for each head. Att i ∈R H×W×p , Att i is the attention weight matrix of the ith head, d k is a scaling factor equal to the dimension of the key vector, Att i The key V applied to the pooling i ′, is the weighted output of the i-th head, and performs residual connection and convolution operations to further capture local feature information. Conv 1×1 () represents a 1×1 convolution, which concatenates the outputs of all heads, focuses on the features of different subspaces, and restores the original channel dimension. X MultiAtt ∈R H×W×C , the concatenated output features are processed through depth-wise separable convolution to reduce the number of parameters and improve computational efficiency, X fused =DSConv(X MultiAtt ), X fused ∈R H ×W×C Finally, the input features are added to the processed features through residual connection, retaining the original information of the input features to obtain the final output, X AttendOutput =X Attin +X fused , X Attendoutput ∈R H×W×C ; S3, construct a pixel focusing block, capture the local feature information of the feature map through flattening, position embedding, normalization, activation, multi-channel 1D convolution, and perform residual connection, and finally output the feature map; In step S3, a pixel-focused block is constructed and the feature X is input Pixelin , X Pixelin ∈R H×W×C , flatten the input features, transform spatial information into sequence information, and use the sequence model's sensitivity to position to capture information, X flat =Flatten(X Pixelin ), X flat ∈R L×C , L is H×W, adding position embedding to the flattened features to enhance position awareness and avoid losing spatial structure of serialized features after flattening, X PE =PE(L)+X flat , X PE ∈R L×C , PE(L) represents a sequence position embedding of length L, and adjusts the feature distribution by normalization to improve the stability of the model, X norm =Norm(X PE ), X norm ∈R L×C , Norm() represents the normalization operation, which is activated by the LeakyReLU function to introduce nonlinearity in the serialized features, so as to better express complex features, X leakyrelu =LeakyReLU(X norm ), X leakyrelu ∈R L×C , LeakyReLU() represents the activation function, which extracts features in the sequence dimension on multiple feature channels through multi-channel 1D convolution, captures the correlation between features at adjacent positions, and performs residual connection to avoid gradient disappearance, X Pixelout =X Pixelin +Concat(MCConv1D(X leakyrelu )), X Pixelout ∈R H×W×C , MCConv1D() represents multi-channel 1D convolution, Concat() represents the concatenation operation, and + represents element-by-element addition; S4. Construct a feature transfer feed-forward network, including global average pooling, two fully connected layers, broadcast context information, and perform residual connection and output; In step S4, a feature transfer feedforward network is constructed, and feature X is input fnin , X fnin ∈R H×W×C , the features are pooled globally to retain the overall features of each channel, X gap =Gap(X fnin ), X gap ∈R 1×1×C , Gap() represents global average pooling, generates adaptive channel weights through two fully connected layers, increases nonlinear expression, and adaptively adjusts the weights of each channel, X mlp =W2·(SiLU(W1·X gap +b1))+b2,X mlp ∈R 1×1×C , W1 and W2 are weight matrices, r is the scaling ratio, b1 and b2 represent bias terms, SiLU() represents the activation function, and the generated weights are broadcast to the spatial dimension H×W. The broadcast weights are multiplied channel by channel with the input features so that the weight of each channel affects the entire feature, and residual connections are performed to avoid gradient disappearance to obtain the output X fnout , X fnout ∈R H×W×c , represents channel-by-channel multiplication, and + represents channel-by-channel element-by-element addition; S5. Construct an efficient feature extraction block, including N efficient feature extraction modules, which are composed of multi-head self-attention, pixel focusing block and feature transfer feedforward network; S6. Build an efficient feature enhancement module, including 1×1 convolution, 5×5 convolution, and 5 efficient feature extraction blocks, and output the enhanced feature map; S7, constructing an ALC green environmentally friendly wallboard defect recognition model, including input, efficient feature enhancement module, ALC green environmentally friendly wallboard defect detection head and output; S8. Train the ALC green environmentally friendly wall panel defect recognition model and detect it. Use the ALC green environmentally friendly wall panel defect data set to train the ALC green environmentally friendly wall panel defect recognition model. After the training is completed, input the new ALC green environmentally friendly wall panel photo to be detected into the ALC green environmentally friendly wall panel defect recognition model to obtain the detection result. The detection result includes the defect location and type in the ALC green environmentally friendly wall panel picture.

2. The ALC green environmental protection wallboard defect identification method according to claim 1 is characterized in that: In step S1, the ALC green environmentally friendly wallboard defect dataset is produced. ALC green environmentally friendly wallboard images containing various defect types are collected, including cracks, bubbles, discoloration, peeling, stains and dents. Each collected image is annotated with defects including the type and location of the defect. The annotated images and corresponding annotation information are organized into a unified format to construct the ALC green environmentally friendly wallboard defect dataset.

3. The ALC green environmental protection wallboard defect identification method according to claim 1 is characterized in that: In step S5, an efficient feature extraction block is constructed and the feature X is input ebin , X ebin ∈R H×W×C , stacking N efficient feature extraction modules, integrating long-distance dependent global features and fine-grained local features, improving the performance and efficiency of the model, X ebout =EFEM N (…EFEM1(X ebin )), EFEM i () is the i-th efficient feature extraction module, i = 1, 2, ..., N, X ebout ∈R H×W×C , i represents the i-th efficient feature extraction module, EFEM(x) = FTFN(MHSA(x) + PFB(x)), MHSA() represents multi-head self-attention, PFB() represents pixel focusing block, and FTFN() represents feature transfer feedforward network.

4. The ALC green environmental protection wallboard defect identification method according to claim 1 is characterized in that: In step S6, an efficient feature enhancement module is constructed and input feature X fein , X fein ∈R H×W×C , obtained by capturing local information through the 5×5 convolution enhancement model and the first efficient feature extraction block By downsampling to reduce the spatial size and computational complexity, the deep features are further extracted through the second efficient feature extraction block. The spatial resolution is further reduced by downsampling to extract deeper features, and the third efficient feature extraction block is used to further extract deep features. Upsampling is used to retain feature information between different resolution layers, residual connections are used to retain the middle-level feature expression, and then 1×1 convolution is used to fuse features. The fourth efficient feature extraction block is used to further extract deep features. The original resolution is restored by upsampling, the residual connection retains the initial feature information, the 1×1 convolution fusion feature and the fifth efficient feature extraction block extract the final enhanced features. Finally, a 5×5 convolution is performed on the final feature map to further smooth and integrate the output X feout , X feout ∈R H×W×C , Conv 5×5 () represents a 5×5 convolution operation.

5. The ALC green environmental protection wallboard defect identification method according to claim 1 is characterized in that: In step S7, the ALC green wallboard defect recognition model is constructed, the image I is input, and the image feature X is obtained through the backbone network. fein , input the efficient feature enhancement module to obtain the enhanced output feature X feout , and then use the detection head to detect and obtain the ALC green environmentally friendly wall panel defect detection results, which include the location and type of the ALC green environmentally friendly wall panel defects.

Images