An image recognition method for curtain wall metal hanging piece detection

Abstract

The application provides an image recognition method for curtain wall metal hanging piece detection, comprising: curtain wall hanging piece data acquisition and preprocessing; based on the data after curtain wall hanging piece data acquisition and preprocessing, a shock reduction super-resolution adversarial neural network is constructed; a classification neural network model is established based on the shock reduction super-resolution adversarial neural network; the classification neural network model is trained and optimized; the classification neural network model is evaluated and tested, and the classification neural network model that passes the test is used to recognize curtain wall hanging piece detection images. The application provides an image recognition method for curtain wall metal hanging piece detection based on a deep learning method, high-resolution images of curtain wall hanging pieces are acquired by adopting a millimeter wave image sensor, then feature extraction and classification are carried out through a deep convolutional neural network, and finally accurate recognition of five different types of curtain wall metal hanging pieces is realized.

Description

Technical Field

[0001] This invention relates to the field of curtain wall inspection technology, and more specifically to an image recognition method for inspecting metal curtain wall fixtures. Background Technology

[0002] Existing image recognition methods include: methods based on traditional machine learning algorithms, methods based on physical methods, and methods based on wavelet transform;

[0003] Methods based on traditional machine learning algorithms, such as Support Vector Machine (SVM) and Random Forest, rely on manually designing features and classifiers, requiring human intervention in data processing and feature selection, and the model accuracy is difficult to improve further.

[0004] Physically based methods: These methods are based on physical principles, such as the interaction between electromagnetic waves and objects. They use numerical calculation methods to simulate or reverse-engineer the electromagnetic properties of target objects, thereby achieving target recognition and detection. Physical methods have good theoretical explanations and mathematical analysis, but they are limited by the complexity of the scene and the large amount of computation in practical applications.

[0005] Wavelet transform-based methods: This method decomposes an image into wavelet coefficients of different frequency bands. By selecting appropriate features and using a classifier for classification, the classification accuracy and robustness of the image can be effectively improved. Commonly used wavelet transform methods include Discrete Wavelet Transform (DWT) and Continuous Wavelet Transform (CWT).

[0006] Methods based on traditional machine learning algorithms have the following problems:

[0007] 1) It requires manual design of features and classifiers, and human involvement in data processing and feature selection; this consumes a lot of time and effort, and in some cases may not be able to find effective features.

[0008] Because traditional machine learning algorithms are sensitive to complex scenes and noise, it is difficult to further improve the accuracy of the models.

[0009] Physically based methods have the following problems:

[0010] 1) Physical methods require basic physical principles and numerical calculation methods, while complex scenarios may make the problem more complicated and difficult, and at the same time require a lot of computing resources.

[0011] 2) Physical methods require predicting the electromagnetic properties of the target object. However, for some irregular metal pendants, their electromagnetic properties are affected by factors such as material and shape, making it difficult to achieve target recognition and detection.

[0012] The wavelet transform-based method has the following problems:

[0013] 1) Wavelet transform is easily affected by image sharpness. In cases where image information is missing or there is too much noise, feature extraction may fail.

[0014] 2) In practical applications, more complex wavelet transform algorithms usually require a lot of computing resources and storage space.

[0015] Therefore, existing technologies have shortcomings and need further improvement. Summary of the Invention

[0016] To address the problems existing in the prior art, this invention provides an image recognition method for detecting metal curtain wall fixtures.

[0017] To achieve the above objectives, the specific solution of the present invention is as follows:

[0018] This invention provides an image recognition method for detecting metal curtain wall fixtures, the method comprising the following steps:

[0019] S1, Data acquisition and preprocessing of curtain wall fixtures;

[0020] S2, based on the data collected and preprocessed from the curtain wall fixtures, constructs a vibration reduction super-resolution adversarial neural network;

[0021] S3, A classification neural network model is established based on a vibration-reducing super-resolution adversarial neural network;

[0022] S4, train and optimize the classification neural network model;

[0023] S5 evaluates and tests the classification neural network model, and uses the tested classification neural network model to identify the detected images of curtain wall hangings.

[0024] Further, step S1 specifically includes: using a pre-labeled five-class classification dataset for curtain wall fixtures, including five types of fixtures: SE, butterfly code, T-shaped fixture, corner code, and back bolt; each type of fixture in the dataset contains 2000 color images of size 256x256, which are divided into training set, validation set, and test set in an 8:1:1 ratio; of which 1600 images are used for training, 200 images are used for validation, and 200 images are used for testing.

[0025] Furthermore, the specific process for constructing the vibration-reducing super-resolution adversarial neural network in step S2 is as follows:

[0026] S201: Use a detection sensor to acquire a low-quality original image A of the curtain wall under simulated vibration conditions, and acquire a high-quality image B of the curtain wall under non-vibration static conditions when the vibration disappears, as the output of the network to be learned.

[0027] S202, image A is downsampled to half its original resolution and used as network input; by training a vibration-reducing super-resolution adversarial neural network, the process of converting low-quality curtain wall image A into high-quality curtain wall image B is realized.

[0028] Furthermore, in step S202, the vibration-reducing super-resolution adversarial neural network is as follows:

[0029]

[0030] G: Generator, which takes a random vector (often called noise) as input and attempts to generate realistic sample data; the goal of the generator is to make the samples it generates mistakenly be considered real data by the discriminator.

[0031] D: Discriminator, which takes real sample data and samples generated by the generator as input and attempts to distinguish between real data and generated data; the goal of the discriminator is to accurately distinguish between real data and generated data;

[0032] E(x): represents the expected value, where x is a variable;

[0033] p_data(x): represents the probability density function of the real data distribution, which describes the statistical characteristics of the real data; p_z(z): represents the probability density function of the random noise input to the generator, which is usually chosen as a known simple distribution, such as Gaussian distribution or uniform distribution;

[0034] D(x): The output of the discriminator to the real data, representing the probability that a given input x is real data; when the input x is a sample drawn from the real data distribution, D(x) should be close to 1;

[0035] D(G(z)): The output of the discriminator to the samples generated by the generator, representing the probability that the data G(z) generated by the given generator is real data; when the data generated by the generator and the real data cannot be distinguished, D(G(z)) should be close to 0;

[0036] The generator G is responsible for generating high-resolution images to reduce the value of V(D,G);

[0037] The generator D is responsible for distinguishing between the generated image and the real high-quality image, and is used to increase the value of V(D,G); therefore, both sides will fall into Nash equilibrium;

[0038] Given G, differentiating D gives:

[0039]

[0040] To obtain the maximum value of this integral, we can find:

[0041]

[0042] By taking argminGmaxDV(G,D)), we can minimize the difference between the two distributions and generate a distribution that is as close as possible to the original distribution, thereby achieving vibration reduction.

[0043] Further, step S3, establishing the classification neural network model, specifically includes:

[0044] S301 introduces the generator part of the deviator super-resolution adversarial neural network into the classification convolutional neural network model;

[0045] S302, input the original vibration small image into the generator to obtain high-quality curtain wall image features under non-vibration static conditions;

[0046] S303 utilizes pixel recombination technology to reduce the resolution of high-quality curtain wall image features by half, while increasing the number of channels to four times the original. It then fuses these features with the features of the original vibration small image.

[0047] S304 inputs the fused features into the classification layer of the network and outputs the classification result.

[0048] Furthermore, step S301 also includes:

[0049] S3011 is an improvement on the standard ResNet classification convolutional neural network model;

[0050] S3012, designed for curtain walls, is called CWNetBlock and is generally referred to as CWNet. It uses depthwise separable convolution and an inverted bottleneck structure.

[0051] S3013 features a separately designed downsampling layer that uses a 2x2 convolution with stride=2 instead of the traditional combination of 3x3 and 1x1 convolutions.

[0052] S3014 adds layer normalization before each downsampling layer to stabilize the training process.

[0053] Further, in step S3014, layer normalization normalizes all neurons in an intermediate layer, setting the net input of the l-th layer neurons to z. (l) Its mean and variance are:

[0054]

[0055]

[0056] z (l) : The net input of neurons in layer l; for a layer with n neurons, z (l) It can be a vector of size n;

[0057] μ (l) In layer normalization, μ (l) This represents the net input z to the neurons in layer l. (l) The mean obtained by averaging along the feature dimension (usually the sample dimension);

[0058] σ (l) In layer normalization, σ (l) This represents the net input z to the neurons in layer l. (l) Calculate the variance along the feature dimension (usually the sample dimension) and take the square root to obtain the standard deviation;

[0059] Where n (l) This represents the number of neurons in the l-th layer.

[0060] Layer normalization is defined as:

[0061]

[0062] γ and β represent the parameter vectors for scaling and translation, and z (l) They have the same dimension.

[0063] Furthermore, step S4 specifically includes: training the network using experimental data, and employing data augmentation, Dropout, and regularization to improve the robustness and accuracy of the model, as well as using an adaptive learning rate to accelerate the training and optimization of the network.

[0064] Further, step S5 specifically includes: evaluating the classification accuracy and recall of the proposed model using a test dataset, and comparing it with traditional machine learning methods.

[0065] The technical solution of this invention has the following beneficial effects:

[0066] 1. Achieve reliable testing of different types of curtain wall metal fittings, with a classification accuracy rate of over 95%;

[0067] 2. Improve the efficiency and accuracy of inspection of curtain wall metal fittings, shorten working time, and improve construction quality and safety;

[0068] 3. By using advanced deep learning algorithms, the automation level of curtain wall metal component inspection is improved, reducing manual intervention and errors;

[0069] 4. It can significantly reduce the cost of metal pendant inspection and improve the economic benefits of enterprises;

[0070] 5. It has a wide range of applications and market prospects, and can be used in fields such as construction, bridges, and transportation;

[0071] 6. It can achieve a high classification accuracy rate, distinguishing five different curtain wall metal hangers. The accuracy rate can reach over 95% in different scenarios, which has high practical value and application prospects. Attached Figure Description

[0072] Figure 1 This is the overall flowchart of the present invention;

[0073] Figure 2 This is a schematic diagram of the vibration-reducing super-resolution adversarial neural network structure of the present invention;

[0074] Figure 3 This is a schematic diagram of the network structure of the present invention;

[0075] Figure 4 This is a schematic diagram of the CWNetBlock structure of the present invention. Detailed Implementation

[0076] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0077] In the description of this invention, unless otherwise explicitly specified and limited, the terms "connected," "linked," and "fixed" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0078] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can include direct contact between the first and second features, or contact between the first and second features through another feature between them. Furthermore, "above," "over," and "on top" of the second feature includes the first feature directly above or diagonally above the second feature, or simply indicates that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature includes the first feature directly below or diagonally below the second feature, or simply indicates that the first feature is at a lower horizontal level than the second feature.

[0079] In the description of this embodiment, the terms "upper," "lower," "front," "rear," "left," and "right," etc., refer to the orientation or positional relationship shown in the accompanying drawings. They are used only for ease of description and simplification of operation, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the present invention. Furthermore, the terms "first" and "second" are used only for descriptive distinction and have no special meaning.

[0080] like Figure 1 As shown, the present invention provides an image recognition method for detecting metal curtain wall fixtures, the method comprising the following steps:

[0081] S1, Data acquisition and preprocessing of curtain wall fixtures;

[0082] Imaging utilizes the propagation characteristics of millimeter waves in materials and objects, which have transmission, reflection, and scattering properties, thus performing well in detecting targets such as metal pendants; this technology can detect different sides of a target by selecting different frequency bands, for example, using 35GHz millimeter waves can be used for surface detection of a target.

[0083] By employing a millimeter-wave image sensor, non-contact scanning and detection of curtain wall metal fixtures are performed to obtain high-resolution images of the fixtures. Then, image processing methods are used to augment the acquired images to improve the accuracy of subsequent classification and reduce network load.

[0084] Data augmentation mainly includes the following methods: RandomRotation, VerticalFlip, HorizontalFlip, Mixup, and CutMix.

[0085] RandomRotation is a random angle rotation.

[0086] VerticalFlip is a vertical mirror flip.

[0087] HorizontalFlip is a horizontal mirror flip.

[0088] The following introduces Mixup data augmentation.

[0089] Suppose there are two images (x) j ,y j ) and (x j ,y j )

[0090] Where x represents the image and y is the one-hot label; the core formula of Mixup is shown below, where λ follows a beta distribution, λ∈0,1;

[0091]

[0092]

[0093] x i and x j : Feature vectors of two input samples; usually represented as tensors of input features;

[0094] y i and y j : The labels of the two input samples; they can be one-hot codes or probability distributions for a classification problem;

[0095] ~x: The feature vector of the newly generated sample; obtained by linear interpolation of the features of the two original samples;

[0096] ~y: The label of the newly generated sample; obtained by linear interpolation of the labels of the two original samples.

[0097] λ: Interpolation coefficient; λ is a random number in the range [0, 1] used to control the degree of mixing between the features and labels of the two samples; when λ = 0, the generated new sample is completely composed of x. j and y j Composition; when λ=1, the newly generated sample is entirely composed of x i and y i Composition; the λ value between 0 and 1 represents the degree of mixing between the features and labels of the two samples.

[0098] The following describes CutMix data augmentation.

[0099] Assume x∈R W×H×C Let y and y represent a training image and its label, respectively; CutMix merges two training samples (x, y, and y) to achieve this. A y A ) and (x B , y B) Generate a new training sample The newly generated training samples are used to train the model; this can be expressed by the formula:

[0100]

[0101]

[0102] Where M∈0,1 W×HIt is a binary mask indicating the positions to be deleted and filled in the two images; 1 is a binary code of all 1s; ⊙ indicates element-wise multiplication; similar to Mixup, the parameter λ belongs to the beta distribution Beta(α, α). In CutMix, α is set to 1, and the parameter λ follows a uniform distribution λ ~ U(0, 1).

[0103] S2, based on the data collected and preprocessed from the curtain wall fixtures, constructs a vibration reduction super-resolution adversarial neural network;

[0104] like Figure 2 As shown in the diagram, the vibration reduction super-resolution adversarial neural network structure diagram is used to collect images of the objects to be detected and background images. Five different curtain wall hanging images are classified into a classification dataset. The training set, validation set, and test set are divided into an 8:1:1 ratio. The vibration reduction super-resolution deep learning neural network is trained. The dataset is expanded using the vibration reduction super-resolution deep learning neural network. The expanded data is used to train a five-class neural network for curtain wall images. Images of the objects to be detected and background images are collected.

[0105] like Figure 3 The vibration-reducing super-resolution adversarial neural network shown mainly involves constructing a vibration-reducing super-resolution adversarial neural network (DeblurGAN) by acquiring a low-quality original image A of the curtain wall under simulated vibration and a high-quality image B of the curtain wall after the vibration has subsided. The specific process is as follows:

[0106] First, a low-quality original image A of the curtain wall is acquired using a detection sensor under simulated vibration conditions. Then, a high-quality image B of the curtain wall is acquired under static conditions where the vibration has disappeared, and used as the output of the network to be learned. Next, image A is downsampled, reducing its resolution to half of its original value, and used as the network input. By training a vibration-reducing super-resolution adversarial neural network, the low-quality curtain wall image A can be transformed into a high-quality curtain wall image B. The innovation lies mainly in the combination of adversarial generative networks and super-resolution reconstruction technology, which can effectively improve the clarity and detail of the curtain wall image. At the same time, the impact of vibration on the quality of the curtain wall image is also considered, and a method for acquiring the original curtain wall image A under simulated vibration conditions is proposed, providing a reliable data foundation for subsequent image processing and analysis.

[0107]

[0108] To provide a loss function for adversarial neural networks,

[0109] G: Generator, which takes a random vector (often called noise) as input and attempts to generate realistic sample data; the goal of the generator is to make the samples it generates mistakenly be considered real data by the discriminator.

[0110] D: Discriminator, which takes real sample data and samples generated by the generator as input and attempts to distinguish between real data and generated data; the goal of the discriminator is to accurately distinguish between real data and generated data;

[0111] E(x): represents the expected value, where x is a variable;

[0112] p_data(x): represents the probability density function of the real data distribution, which describes the statistical characteristics of the real data; p_z(z): represents the probability density function of the random noise input to the generator, which is usually chosen as a known simple distribution, such as Gaussian distribution or uniform distribution;

[0113] D(x): The output of the discriminator to the real data, representing the probability that a given input x is real data; when the input x is a sample drawn from the real data distribution, D(x) should be close to 1;

[0114] D(G(z)): The output of the discriminator to the samples generated by the generator, representing the probability that the data G(z) generated by the given generator is real data; when the data generated by the generator and the real data cannot be distinguished, D(G(z)) should be close to 0;

[0115] Generator G is responsible for generating high-resolution images and minimizing the value of V(D,G) as much as possible; generator D is responsible for distinguishing between the generated images and the real high-resolution images and maximizing the value of V(D,G) as much as possible; therefore, both will fall into Nash equilibrium.

[0116] Given G, differentiating D gives:

[0117]

[0118] To obtain the maximum value of this integral, we can find:

[0119]

[0120] At this point, by taking argminGmaxDV(G,D)), we can find G that minimizes the difference between the two distributions, thus generating a distribution that is as close as possible to the original distribution.

[0121] S3, A classification neural network model is established based on a vibration-reducing super-resolution adversarial neural network;

[0122] For the task of classifying and detecting curtain wall components, this patent introduces a generator part of a "vibration-reducing super-resolution adversarial neural network" into the classification convolutional neural network model. Specifically, the original small vibration image is first input into the generator to obtain high-quality curtain wall image features under non-vibration static conditions. Then, using pixel recombination technology, the resolution of the high-quality curtain wall image features is reduced by half, and the number of channels is quadrupled. These features are then fused with the features of the original small vibration image. Finally, the fused features are input into the classification layer of the network to output the classification result.

[0123] like Figure 4 The diagram shows the CWNet structure. This patent improves upon the standard ResNet in terms of network structure. While retaining the overall ResNet structure, each block is modified. A CWNetBlock is designed specifically for curtain walls, collectively referred to as CWNet, using depthwise separable convolutions and an inverted bottleneck structure. Since frequent nonlinear projections are detrimental to the transmission of network feature information, the number of activation layers is reduced to maintain the transmission of network feature information. This patent also features a separate downsampling layer, employing a 2x2 convolution with stride=2 instead of the traditional combination of 3x3 and 1x1 convolutions. This design aims to avoid training errors. To ensure stability, this patent adds Layer Normalization (LN) before each downsampling layer to stabilize the training process. Experimental results show that the above network model exhibits excellent accuracy and robustness in the curtain wall component classification and detection task. Compared with the traditional ResNet, the CWNet network structure of this patent can extract features better, while having higher computational efficiency and stability. In summary, this patent proposes a vibration-reducing super-resolution adversarial neural network and a CWNet classification convolutional neural network model, which can be used for the five-class classification detection task of curtain wall components, greatly improving the detection accuracy and stability, and has broad application prospects. LN stands for Layer Normalization, which normalizes all neurons in an intermediate layer.

[0124] Let the net input of the l-th layer of nerves be z. (l) Its mean and variance are:

[0125]

[0126]

[0127] z (l) : The net input of neurons in layer l; for a layer with n neurons, z (l) It can be a vector of size n;

[0128] μ(l) In layer normalization, μ (l) This represents the net input z to the neurons in layer l. (l) The mean obtained by averaging along the feature dimension (usually the sample dimension);

[0129] σ (l) In layer normalization, σ (l) This represents the net input z to the neurons in layer l. (l) Calculate the variance along the feature dimension (usually the sample dimension) and take the square root to obtain the standard deviation;

[0130] Where n (l) This represents the number of neurons in the l-th layer.

[0131] Layer normalization is defined as:

[0132]

[0133] γ and β represent the parameter vectors for scaling and translation, and z (l) They have the same dimension.

[0134] S4, train and optimize the classification neural network model;

[0135] To learn more accurate features, the network was trained using a large amount of experimental data, and various techniques were employed to improve the robustness and accuracy of the model, such as data augmentation, Dropout, and regularization. In addition, an adaptive learning rate method was used to accelerate the training and optimization of the network.

[0136] S5 evaluates and tests the classification neural network model, and uses the tested classification neural network model to identify the detected images of curtain wall hangings.

[0137] The proposed model's classification accuracy, recall, and other metrics will be evaluated using a test dataset and compared with traditional machine learning methods. As for the challenges of testing, complex lighting conditions, size variations, and noise are often encountered when inspecting curtain wall components. Therefore, we will address these challenges by implementing appropriate processing and testing methods.

[0138] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. All equivalent structural transformations made under the inventive concept of the present invention using the contents of the present invention specification and drawings, or direct / indirect applications in other related technical fields, are included within the protection scope of the present invention.

Claims

1. An image recognition method for detecting metal curtain wall fixtures, characterized in that, Includes the following steps: S1, Data acquisition and preprocessing of curtain wall fixtures; S2, based on the data collected and preprocessed from the curtain wall fixtures, constructs a vibration reduction super-resolution adversarial neural network; S3, A classification neural network model is established based on a vibration-reducing super-resolution adversarial neural network; S4, train and optimize the classification neural network model; S5 evaluates and tests the classification neural network model, and uses the tested classification neural network model to identify the detected images of curtain wall hangings; The specific process for constructing the vibration-reducing super-resolution adversarial neural network in step S2 is as follows: S201: Use a detection sensor to acquire a low-quality original image A of the curtain wall under simulated vibration conditions, and acquire a high-quality image B of the curtain wall under non-vibration static conditions when the vibration disappears, as the output of the network to be learned. S202, image A is downsampled to half its original resolution and used as network input; by training a vibration-reducing super-resolution adversarial neural network, the process of converting low-quality curtain wall image A into high-quality curtain wall image B is realized. Step S3 involves establishing the classification neural network model, specifically including: S301 introduces the generator part of the deviator super-resolution adversarial neural network into the classification convolutional neural network model; S302, input the original vibration small image into the generator to obtain high-quality curtain wall image features under non-vibration static conditions; S303 utilizes pixel recombination technology to reduce the resolution of high-quality curtain wall image features by half, while increasing the number of channels to four times the original. It then fuses these features with the features of the original small vibration image. S304, the fused features are input into the classification layer of the network, and the classification result is output; Step S301 also includes: S3011 is an improvement on the standard ResNet classification convolutional neural network model; S3012, designed for curtain walls, is called CWNetBlock and is generally referred to as CWNet. It uses depthwise separable convolution and an inverted bottleneck structure. S3013 features a separately designed downsampling layer that uses a 2x2 convolution with stride=2 instead of the traditional combination of 3x3 and 1x1 convolutions. S3014 adds layer normalization before each downsampling layer to stabilize the training process.

2. The image recognition method for detecting metal curtain wall fixtures according to claim 1, characterized in that, Step S1 specifically includes: using a pre-labeled five-class dataset of curtain wall fixtures, including five types of fixtures: SE, butterfly code, T-shaped fixture, corner code, and back bolt; each type of fixture in the dataset contains 2000 color images of size 256x256, which are divided into training set, validation set, and test set in an 8:1:1 ratio.

3. The image recognition method for detecting metal curtain wall fixtures according to claim 1, characterized in that, In step S202, the vibration reduction super-resolution adversarial neural network is as follows: G: Generator, which takes a random vector as input and attempts to generate realistic sample data; the goal of the generator is to make the samples it generates mistakenly be identified as real data by the discriminator. D: Discriminator, which accepts real sample data and samples generated by the generator as input, and distinguishes between real data and generated data; the goal of the discriminator is to accurately distinguish between real data and generated data. E(x): represents the expected value, where x is a variable; p_data(x): represents the probability density function of the real data distribution, which describes the statistical characteristics of the real data; p_z(z): represents the probability density function of the random noise input to the generator, which is chosen to be a known simple distribution; D(x): The output of the discriminator to the real data, representing the probability that a given input x is real data; when the input x is a sample drawn from the real data distribution, D(x) is close to 1; D(G(z)): The output of the discriminator to the samples generated by the generator, representing the probability that the data G(z) generated by the given generator is real data; when the data generated by the generator and the real data cannot be distinguished, D(G(z)) is close to 0; The generator G is responsible for generating high-resolution images to reduce resolution. The value; Generator D is responsible for distinguishing between the generated image and the real high-quality image, and is used to increase... The value of ; therefore, both sides will fall into Nash equilibrium; Given G, differentiating D gives: To obtain the maximum value of this integral, we can find: Pick By choosing G such that the difference between the two distributions is minimized, a distribution that is as close as possible to the original distribution can be generated, thereby achieving vibration reduction.

4. The image recognition method for detecting metal curtain wall fixtures according to claim 1, characterized in that, In step S3014, layer normalization normalizes all neurons in an intermediate layer, letting the first... The net input of the layer nerve is Its mean and variance are: : No. The net input of a layer of neurons; for a given layer of neurons... A layer of neurons, It is a size of ; In layer normalization, Indicates the first Net input of layer neurons The mean obtained by averaging along the feature dimension; In layer normalization, Indicates the first Net input of layer neurons Calculate the variance along the feature dimension and take the square root to obtain the standard deviation; in For the first The number of neurons in the layer; Layer normalization is defined as: , The parameter vectors representing scaling and translation, and They have the same dimension.

5. The image recognition method for detecting metal curtain wall fixtures according to claim 1, characterized in that, Step S4 specifically includes: training the network using experimental data, and using data augmentation, Dropout, and regularization to improve the robustness and accuracy of the model, as well as using an adaptive learning rate to accelerate the training and optimization of the network.

6. The image recognition method for detecting metal curtain wall fixtures according to claim 1, characterized in that, Step S5 specifically includes: evaluating the classification accuracy and recall of the proposed model using a test dataset, and comparing it with traditional machine learning methods.

Images