Image recognition data model based on multi-technology fusion convolutional neural network
By introducing spatial transformation network, deep residual network and hollow convolution into convolution neural networks, the problems of gradient disappearance and insufficient robustness of spatial transformation in image recognition are solved, and more efficient image features are achieved Understand and identify.
Patent Information
- Application Number
- CN202411107078.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-13
- Publication Date
- 2024-11-12
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing image recognition technology based on convolutional neural networks is prone to the problem of gradient disappearance or gradient explosion, and is not very robust to spatial transformations such as translation, rotation, and scaling of images, making it difficult to capture global information.
Multi-technology fusion convolutional neural networks are adopted, including spatial transformation network, deep residual network and hollow convolution, through preprocessing and enhancement layers, spatial transformation network module, deep residual network module, hollow convolution module, feature fusion and classification layer, improve the model's adaptability to spatial transformation and the training ability of deep networks.
It significantly improves the accuracy and robustness of image recognition, can more effectively capture local and global features of the image, solves the problems of gradient vanishing and explosion, and enhances the adaptability to spatial transformation.
Smart Images

Figure FDA0004991130940000031
Abstract
Description
Technical Field
[0001] The present invention relates to the field of computer vision technology, and in particular to an image recognition data model based on a multi-technology fusion convolutional neural network. Background Art
[0002] In industrial production, automated image recognition technology can replace manual labor for heavy and repetitive inspection work, reduce labor costs and improve production efficiency. At the same time, image recognition technology can enable equipment to have the ability to automatically analyze, understand and respond, thereby achieving automation and intelligence. Early image recognition technology only used simple pattern recognition algorithms to identify basic shapes and structures in images. With the improvement of computer performance, traditional machine learning methods based on feature extraction and pattern matching have emerged, which have improved the accuracy of image recognition to a certain extent. With the rise of deep learning and neural network technology, image recognition technology has ushered in a revolutionary change. The emergence of deep learning models, especially convolutional neural networks, has greatly improved the accuracy of image recognition and can handle more complex image features. In addition, the establishment of large-scale data sets and the improvement of computing power have also provided support for the widespread application of deep learning in the field of image recognition.
[0003] In recent years, although image recognition technology has been continuously improved, there are still many technical defects, including at least the following aspects: 1. The currently commonly used image recognition technology based on convolutional neural networks is prone to gradient vanishing or gradient exploding problems, making deep networks difficult to train. 2. Convolutional neural networks are not very robust to spatial transformations such as translation, rotation, and scaling of images, and require more data enhancement and model design to enhance their adaptability to spatial transformations. 3. The convolution kernel of each layer of the convolutional neural network can only capture local features of a certain size, resulting in limited grasp of the overall information of the image. For larger objects or scenes, a larger receptive field may be required to capture global information. Summary of the invention
[0004] The technical problem to be solved by the present invention is to provide an image recognition data model based on a multi-technology fusion convolutional neural network in view of the various deficiencies in the prior art.
[0005] In order to solve the above technical problems, the technical solutions adopted by the present invention are as follows.
[0006] An image recognition data model based on a multi-technology fusion convolutional neural network improves the model's adaptability to spatial transformation and its ability to understand complex image features by establishing a multi-technology fusion convolutional neural network that includes a spatial transformation network, a deep residual network, and a dilated convolution. It also solves the problem of deep networks being difficult to train due to gradient vanishing and gradient exploding.
[0007] As a preferred technical solution of the present invention, the model includes a preprocessing and enhancement layer, a spatial transformation network module, a deep residual network module, a hole convolution module, a feature fusion and classification layer.
[0008] As a preferred technical solution of the present invention, the preprocessing and enhancement layer uses mathematical transformation-based methods including normalization and histogram equalization to preprocess and enhance the input image to ensure the consistency of data distribution and improve the contrast of the image, making it easier for the model to recognize image features: a. Standardize the input image to ensure that the mean of the image data is 0 and the variance is 1:
[0009] X_norm=(X-μ) / σ
[0010] Where X is the original image data, μ and σ are the mean and standard deviation of all image pixel values in the data set respectively;
[0011] b. Use data augmentation techniques including rotation, scaling, and cropping to increase the generalization ability of the model.
[0012] As a preferred technical solution of the present invention, the spatial transformation network module uses a learnable affine transformation to correct the input image and enhance the invariance of the model to the spatial transformation of the image. The specific mathematical structure is as follows:
[0013] a. Set the input image to be represented as I, and preprocess I to make it meet the network input requirements;
[0014] b. Use a small CNN network, the output is 6 parameters, forming the affine transformation matrix θ:
[0015] θ=[θ 11 θ 12 θ 13 ]
[0016] [θ 21 θ 22 θ 23 ]
[0017] where each θ ij It is obtained through online learning;
[0018] c. For each pixel position (x s ,y s ), and use θ to calculate its new position (x t ,y t ):
[0019] [x t ][θ 11 θ 12 θ 13 ]
[0020] [y t ]=[θ 21 θ 22 θ 23 ]
[0021] d. Generate the transformed image using the new coordinate grid. For each transformed coordinate (x t ,y t ), use bilinear interpolation to sample pixel values from the original image I:
[0022] I'(x t ,y t ))=∑{i,j}I(i,j)*max(0,1-|x t -i|)*max(0,1-|y t -j|)
[0023] where (i,j) is the pixel position in I, and I'(x t ,y t ) is the transformed image at (x t ,y t )’s pixel value;
[0024] The transformed image I' has a rectified spatial structure and can be fed into subsequent network layers for further feature extraction and classification.
[0025] As a preferred technical solution of the present invention, the deep residual network module adopts residual connection to prevent the gradient vanishing problem and deepen the network depth without losing the training ability. The specific mathematical structure is as follows:
[0026] The residual module is designed for the preprocessed and enhanced image, and each basic residual module is expressed as:
[0027] F(x)=ReLU(x+Conv(ReLU(Conv(x))))
[0028] Among them, x is the input feature map, Conv represents the convolution operation, and ReLU is the activation function, so as to achieve fast feature transmission;
[0029] In the deep residual network, the above residual modules are repeatedly stacked to build a deep network structure. In order to prevent the gradient vanishing problem, skip connections are used to add the input directly to the subsequent layers.
[0030] As a preferred technical solution of the present invention, the dilated convolution module introduces a dilated convolution layer in the deep layer to expand the receptive field and capture a wider range of contextual information without adding additional parameters. The specific mathematical structure is as follows:
[0031] The atrous convolution is designed for the preprocessed and enhanced image. The receptive field is expanded by increasing the step size of the convolution kernel between the standard convolution kernel elements. For a given input feature map x, the output calculation formula of the atrous convolution is:
[0032] F(x)=x*W_dilated
[0033] Among them, * represents the convolution operation, W_dilated represents the expanded convolution kernel, which is achieved by increasing the spatial interval between the elements of the standard convolution kernel. In the deep network, the standard convolution layer and the hole convolution layer design are used alternately. The standard convolution layer is used to capture local features, and the hole convolution layer is used to expand the receptive field and capture a wider range of contextual information.
[0034] As a preferred technical solution of the present invention, the feature fusion and classification layer uses a pooling layer to reduce the spatial dimension of the feature, integrates the multi-scale feature maps through a fusion layer, and then classifies the features through a fully connected layer.
[0035] As a preferred technical solution of the present invention, after the model is built, the cross entropy loss function is used to evaluate the performance of the model to reduce the overfitting phenomenon. The formula is:
[0036] L=-Σ{c=1}^My_{o,c}log(p_{o,c})
[0037] Where M is the number of classes, y_{o,c} is 1 if the sample belongs to class c and 0 otherwise, and p_{o,c} is the predicted probability that the sample belongs to class c.
[0038] As a preferred technical solution of the present invention, after the model is built, the Adam optimizer is used to update the network parameters in combination with the gradient descent method:
[0039]
[0040] Among them, θ represents the network parameters, η is the learning rate, is the gradient of the loss function with respect to θ; a learning rate decay strategy is adopted to gradually reduce the learning rate as the training progresses to refine the adjustment of network parameters; a weight decay term is added to the loss function to prevent the model from overfitting and enhance the generalization ability of the model; batch normalization is used after each convolutional layer to speed up the training process and improve model stability.
[0041] As a preferred technical solution of the present invention, after the model is constructed, the softmax function is used as the activation function to generate the classification probability in the output layer, softmax(z_i)=exp(z_i-max(z)) / Σ_{j=1}^Kexp(z_j), where z_i is the i-th element of the network output layer, K is the total number of categories, exp is an exponential function, and the output is a vector of K, each element of which represents the predicted probability of the corresponding category.
[0042] The beneficial effect of adopting the above technical solution is that: based on the original convolutional neural network graphic recognition technology, the present invention introduces a spatial transformation network to explicitly perform spatial transformation on the input image, enhance the model's adaptability to spatial transformation, and improve the robustness to image transformation. By introducing residual connections, information can be directly propagated through identity mapping to alleviate the problems of gradient vanishing and explosion, and improve the depth of the network without losing training ability. By combining the advantages of standard convolution and hole convolution, the local features of the image and a wide range of contextual information are effectively captured without adding additional model parameters, thereby improving the efficiency and effect of the model. Integrating the above technology into a convolutional neural network solves the various defects of the current image recognition technology based on convolutional neural networks, making image recognition more accurate and effective.
[0043] The following embodiments describe in detail the technical advantages and beneficial effects of various technical details of the present invention. DETAILED DESCRIPTION
[0044] The following examples describe the present invention in detail. In the description of the following examples, specific details such as specific system structures, technologies, etc. are proposed for illustration rather than limitation, so as to provide a thorough understanding of the present application embodiments. However, it should be clear to those skilled in the art that the present application can also be implemented in other embodiments without these specific details. In other cases, the detailed description of well-known systems, devices, circuits and methods is omitted to prevent unnecessary details from hindering the description of the present application.
[0045] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or their collections. It should also be understood that the term "and / or" used in the present specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes these combinations.
[0046] As used in the specification of this application and the appended claims, the term "if" may be interpreted as "when" or "uponce" or "in response to determining" or "in response to detecting" depending on the context. Similarly, the phrase "if it is determined" or "if [the described condition or event] is detected" may be interpreted as meaning "uponce it is determined" or "in response to determining" or "uponce [the described condition or event] is detected" or "in response to detecting [the described condition or event]" depending on the context. In addition, in the description of the specification of this application and the appended claims, the terms "first", "second", "third", etc. are only used to distinguish the description and cannot be understood as indicating or implying relative importance.
[0047] References to "one embodiment" or "some embodiments" etc. described in the specification of this application mean that one or more embodiments of the present application include specific features, structures or characteristics described in conjunction with the embodiment. Therefore, the statements "in one embodiment", "in some embodiments", "in some other embodiments", "in some other embodiments", etc. that appear in different places in this specification do not necessarily refer to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized in other ways. The terms "including", "comprising", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized in other ways.
[0048] Example 1
[0049] The image recognition data model based on multi-technology fusion convolutional neural network includes preprocessing and enhancement layers, spatial transformation network module, deep residual network module, void convolution module, feature fusion and classification layer.
[0050] The preprocessing and enhancement layers use mathematical transformation-based methods including normalization and histogram equalization to preprocess and enhance the input image to ensure the consistency of data distribution and improve the contrast of the image, making it easier for the model to recognize image features: a. Standardize the input image to ensure that the mean of the image data is 0 and the variance is 1:
[0051] X_norm=(X-μ) / σ
[0052] Where X is the original image data, μ and σ are the mean and standard deviation of all image pixel values in the data set respectively;
[0053] b. Use data augmentation techniques including rotation, scaling, and cropping to increase the generalization ability of the model.
[0054] The feature fusion and classification layer uses a pooling layer to reduce the spatial dimension of the features, integrates the multi-scale feature maps through a fusion layer, and then classifies the features through a fully connected layer.
[0055] Feature fusion is the integration of feature information from different levels or sources to provide a more comprehensive and rich feature representation. It is usually performed between multiple convolutional layers. Through feature fusion, features of different levels, scales or semantic levels can be effectively combined, thereby improving the model's ability to understand images and recognition accuracy.
[0056] The pooling layer is used to reduce the number of model parameters, control overfitting and improve computational efficiency. The maximum value of each region of the feature map is taken as the output through maximum pooling, and the average value of each region of the feature map is taken as the output through average pooling. Through the pooling operation, the size of the feature map is reduced, but the main feature information is retained, which helps to improve the robustness and generalization ability of the model.
[0057] The fully connected layer is used to classify features. Through the fully connected layer, the model can map the high-level features extracted and fused into the category space to make classification decisions.
[0058] Finally, the activation function is used through the output layer to generate classification probabilities, thus completing the image recognition task.
[0059] After the model is built, the cross entropy loss function is used to evaluate the performance of the model and reduce overfitting. The formula is:
[0060] L=-Σ{c=1}^My_{o,c}log(p_{o,c})
[0061] Where M is the number of classes, y_{o,c} is 1 if the sample belongs to class c and 0 otherwise, and p_{o,c} is the predicted probability that the sample belongs to class c.
[0062] The cross entropy loss function measures the difference between the model's prediction of the correct category and the actual label during the loss calculation process, making the model more inclined to produce more accurate predictions during the training process, that is, assigning a higher probability to the correct category, which helps prevent the model from overfitting the training data during the training process, thereby reducing the overfitting phenomenon.
[0063] The calculation of the cross entropy loss function involves taking the logarithm of the predicted probability and then multiplying it by the actual label value. This process can help alleviate the problem of gradient explosion, which can cause model parameter updates to be too large in deep learning models, making the model unable to converge or causing overfitting. By taking the logarithm of the predicted probability, the gradient value is limited, making the gradient update more stable and helping to alleviate overfitting.
[0064] The cross entropy loss function provides a good gradient signal for classification tasks, which helps to train deep learning models. Compared with other loss functions, the cross entropy loss function is easier to optimize in classification tasks, and its gradient is clearer and easier to propagate. This makes it easier for the model to learn the characteristics of the data and less prone to overfitting.
[0065] After the model is built, the Adam optimizer is used to update the network parameters in combination with the gradient descent method:
[0066]
[0067] Among them, θ represents the network parameters, η is the learning rate, is the gradient of the loss function with respect to θ; a learning rate decay strategy is adopted to gradually reduce the learning rate as the training progresses to refine the adjustment of network parameters; a weight decay term is added to the loss function to prevent the model from overfitting and enhance the generalization ability of the model; batch normalization is used after each convolutional layer to speed up the training process and improve model stability.
[0068] After the model is built, the softmax function is used as the activation function to generate the classification probability in the output layer:
[0069] softmax(z_i)=exp(z_i-max(z)) / Σ_{j=1}^Kexp(z_j), where z_i is the i-th element of the network output layer, K is the total number of categories, exp is an exponential function, and the output is a vector of K, each element representing the predicted probability of the corresponding category. In the image recognition data model, the main task of the output layer is to convert the features learned by the network into the predicted probability of the category to which the image belongs, a vector containing the probabilities of each category, whose length is equal to the number of categories in the data set. For example, if it is a 10-category classification task, the vector output by the output layer will contain 10 elements, each representing the predicted probability of the corresponding category. These probability values are normalized by the Softmax function to ensure that they are all between 0 and 1 and their sum is 1.
[0070] Example 2
[0071] The spatial transformation network module uses learnable affine transformation to correct the input image and enhance the model's invariance to spatial transformation of the image. The specific mathematical structure is as follows:
[0072] First, the input image is represented as I, and I is preprocessed to meet the network input requirements;
[0073] Secondly, a small convolutional neural network is used to process the image and output 6 parameters to form the affine transformation matrix θ. This matrix contains the information of rotation, translation and scaling required for the transformation, and each parameter θij is learned through the network.
[0074] θ=[θ 11 θ 12 θ 13 ]
[0075] [θ 21 θ 22 θ 23 ]
[0076] Third, for each pixel position (x s ,y s ), and use θ to calculate its new position (x t ,y t ):
[0077] [x t ][θ 11 θ 12 θ 13 ]
[0078] [y t ]=[θ 21 θ 22 θ 23 ]
[0079] Fourth, use the new coordinate grid to generate the transformed image. For each transformed coordinate (x t ,y t ), bilinear interpolation is used to sample pixel values from the original image I to preserve the image details and ensure the quality of the transformed image:
[0080] I'(x t ,y t ))=∑{i,j}I(i,j)*max(0,1-|x t -i|)*max(0,1-|y t -j|)
[0081] where (i,j) is the pixel position in I, and I'(x t ,y t ) is the transformed image at (x t ,y t )’s pixel value;
[0082] After the above steps, the corrected image I' can be obtained. The transformed image I' has a corrected spatial structure and can be sent to the subsequent network layer for further feature extraction and classification.
[0083] Example 3
[0084] The deep residual network module uses residual connections to prevent the gradient vanishing problem and deepen the network depth without losing training ability. The specific mathematical structure is as follows:
[0085] First, the input images are preprocessed and enhanced to suit the needs of the network and ensure the quality and diversity of image data.
[0086] Secondly, the residual module is designed to achieve fast feature transfer by designing a basic residual module. The core idea of the deep residual network module is to directly connect the input feature map to the subsequent layer, and perform a series of operations between these two points, such as two convolution operations, and a nonlinear activation function is applied after each convolution. This design allows the input feature map to be transmitted not only through the traditional inter-layer transfer method, but also directly transferred to deeper layers through "jump connections", thereby solving the problem of gradient disappearance caused by increasing network depth.
[0087] The residual module is designed for the preprocessed and enhanced image, and each basic residual module is expressed as:
[0088] F(x)=ReLU(x+Conv(ReLU(Conv(x))))
[0089] Among them, x is the input feature map, Conv represents the convolution operation, and ReLU is the activation function, so as to achieve fast feature transmission;
[0090] Third, the stacking of residual modules. In the deep residual network, the basic residual modules described above are repeatedly stacked to form a deep network structure. This structural design allows the network to be very deep without losing the ability to train, because through the jump connection, the gradient can flow directly to the shallower layer, avoiding the common gradient vanishing problem in deep networks.
[0091] Fourth, in order to prevent the gradient vanishing problem, the residual network deliberately introduces skip connections to add the input feature map directly to the subsequent layers. This mechanism not only helps maintain the integrity of the feature map, but also ensures that information can be effectively transmitted even if the network is very deep.
[0092] Example 4
[0093] First, the input image is preprocessed and enhanced to ensure that the image data is suitable for network processing and has rich and diverse information.
[0094] Secondly, the dilated convolution module introduces a dilated convolution layer in the deep layer to expand the receptive field and capture a wider range of contextual information without adding additional parameters. The specific mathematical structure is as follows: For a given input feature map x, the output calculation formula using the dilated convolution is:
[0095] F(x)=x*W_dilated
[0096] Among them, * represents the convolution operation, and W_dilated represents the expanded convolution kernel.
[0097] Third, in the design of deep networks, by alternating between standard convolutional layers and dilated convolutional layers, accurate capture of local features and comprehensive understanding of extensive contextual information can be achieved. Standard convolutional layers are mainly responsible for extracting local features, while dilated convolutional layers are responsible for capturing broader contextual information through their enlarged receptive fields.
[0098] In each embodiment, the hardware implementation of the technology can directly adopt existing intelligent devices, including but not limited to industrial computers, PCs, smart phones, handheld stand-alone machines, floor-standing stand-alone machines, etc. Its input device preferably adopts a screen keyboard, its data storage and calculation module adopts existing memory, calculator, controller, its internal communication module adopts existing communication ports and protocols, and its remote communication adopts existing GPRS networks, World Wide Internet, etc. It can be clearly understood by technicians in the relevant field that for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example. In actual applications, the above-mentioned function allocation can be completed by different functional units and modules as needed, that is, the internal structure of the device is divided into different functional units or modules to complete all or part of the functions described above. Each functional unit and module in the embodiment can be integrated in a processing unit, or each unit can exist physically separately, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of each functional unit and module are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the above-mentioned method embodiment, and will not be repeated here. In the embodiments provided by the present invention, it should be understood that the disclosed device / terminal equipment and method can be implemented in other ways. For example, the device / terminal equipment embodiments described above are only schematic, for example, the division of modules or units is only a logical function division, and there may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point, the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms. The unit described as a separate component may or may not be physically separated, and the component displayed as a unit may or may not be a physical unit, that is, it may be located in one place, or it may be distributed on multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the scheme of this embodiment. The functional units in various embodiments of the present invention can be integrated in a processing unit, or each unit can exist physically alone, or two or more units can be integrated in one unit.
[0099] The embodiments described above are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included in the protection scope of the present invention.
Claims
1. An image recognition data model based on a multi-technology fusion convolutional neural network, characterized by: By establishing a multi-technology fusion convolutional neural network including spatial transformation network, deep residual network and void convolution, the model's adaptability to spatial transformation is improved, the ability to understand complex image features is enhanced, and the problem of deep network difficulty in training caused by gradient disappearance and gradient explosion is solved.
2. The image recognition data model based on multi-technology fusion convolutional neural network according to claim 1 is characterized in that: The model contains preprocessing and enhancement layers, spatial transformer network modules, deep residual network modules, dilated convolution modules, feature fusion and classification layers.
3. The image recognition data model based on multi-technology fusion convolutional neural network according to claim 2 is characterized in that: The preprocessing and enhancement layers use mathematical transformation-based methods including normalization and histogram equalization to preprocess and enhance the input image to ensure the consistency of data distribution and improve the contrast of the image, making it easier for the model to recognize image features: a. Standardize the input image to ensure that the mean of the image data is 0 and the variance is 1: X_norm=(X-μ) / σ Where X is the original image data, μ and σ are the mean and standard deviation of all image pixel values in the data set respectively; b. Use data augmentation techniques including rotation, scaling, and cropping to increase the generalization ability of the model.
4. The image recognition data model based on multi-technology fusion convolutional neural network according to claim 2 is characterized in that: The spatial transformation network module uses learnable affine transformation to correct the input image and enhance the model's invariance to spatial transformation of the image. The specific mathematical structure is as follows: a. Set the input image to be represented as I, and preprocess I to make it meet the network input requirements; b. Use a small CNN network, the output is 6 parameters, forming the affine transformation matrix θ: θ=[θ 11 i 12 i 13 ] [i 21 i 22 i 23 ] where each θ ij It is obtained through online learning; c. For each pixel position (x s ,y s ), and use θ to calculate its new position (x t ,y t ): [x] t ][θ 11 i 12 i 13 ] [y t ]=[θ 21 i 22 i 23 ] d. Generate the transformed image using the new coordinate grid. For each transformed coordinate (x t ,y t ), use bilinear interpolation to sample pixel values from the original image I: I'(x t ,y t ))=∑{i,j}I(i,j)*max(0,1-|x t -i|)*max(0,1-|y t -j|) where x t ,y t is the coordinate of the transformed image, i and j are the coordinates of the original image, I(i,j) is the pixel value of the original image at the coordinate (i,j), and I'(x t ,y t ) is the transformed image at (x t ,y t ), max(0,1-|x t -i))、max(0,1-|y t -j|) is the weight of bilinear interpolation, which is used to calculate the value of each pixel in the output image; The transformed image I' has a rectified spatial structure and can be fed into subsequent network layers for further feature extraction and classification.
5. The image recognition data model based on multi-technology fusion convolutional neural network according to claim 2 is characterized in that: The deep residual network module uses residual connections to prevent the gradient vanishing problem and deepen the network depth without losing training ability. The specific mathematical structure is as follows: The residual module is designed for the preprocessed and enhanced image, and each basic residual module is expressed as: F(x)=ReLU(x+Conv(ReLU(Conv(x)))) Among them, x is the input feature map, Conv represents the convolution operation, and ReLU is the activation function, so as to achieve fast feature transmission; In the deep residual network, the above residual modules are repeatedly stacked to build a deep network structure. In order to prevent the gradient vanishing problem, skip connections are used to add the input directly to the subsequent layers.
6. The image recognition data model based on multi-technology fusion convolutional neural network according to claim 2 is characterized in that: The dilated convolution module introduces a dilated convolution layer in the deep layer to expand the receptive field and capture a wider range of contextual information without adding additional parameters. The specific mathematical structure is as follows: The atrous convolution is designed for the preprocessed and enhanced image. The receptive field is expanded by increasing the step size of the convolution kernel between the standard convolution kernel elements. For a given input feature map x, the output calculation formula of the atrous convolution is: F(x)=x*W_dilated Among them, * represents the convolution operation, W_dilated represents the expanded convolution kernel, which is achieved by increasing the spatial interval between the elements of the standard convolution kernel. In the deep network, the standard convolution layer and the hole convolution layer design are used alternately. The standard convolution layer is used to capture local features, and the hole convolution layer is used to expand the receptive field and capture a wider range of contextual information.
7. The image recognition data model based on multi-technology fusion convolutional neural network according to claim 2 is characterized in that: The feature fusion and classification layer uses a pooling layer to reduce the spatial dimension of the features, integrates the multi-scale feature maps through a fusion layer, and then classifies the features through a fully connected layer.
8. The image recognition data model based on multi-technology fusion convolutional neural network according to any one of claims 1 to 7, characterized in that: After the model is built, the cross entropy loss function is used to evaluate the performance of the model and reduce overfitting. The formula is: L=-Σ{c=1}^My_{o,c}log(p_{o,c}) Where M is the number of classes, y_{o,c} is 1 if the sample belongs to class c and 0 otherwise, and p_{o,c} is the predicted probability that the sample belongs to class c.
9. The image recognition data model based on multi-technology fusion convolutional neural network according to any one of claims 1 to 7, characterized in that: After the model is built, the Adam optimizer is used to update the network parameters in combination with the gradient descent method: Among them, θ represents the network parameters, η is the learning rate, is the gradient of the loss function with respect to θ; a learning rate decay strategy is adopted to gradually reduce the learning rate as the training progresses to refine the adjustment of network parameters; a weight decay term is added to the loss function to prevent the model from overfitting and enhance the generalization ability of the model; batch normalization is used after each convolutional layer to speed up the training process and improve model stability.
10. The image recognition data model based on multi-technology fusion convolutional neural network according to any one of claims 1 to 7, characterized in that: After the model is built, the softmax function is used as the activation function to generate the classification probability in the output layer: softmax(z_i)=exp(z_i-max(z)) / Σ_{j=1}^Kexp(z_j), where z_i is the i-th element of the network output layer, K is the total number of categories, exp is an exponential function, and the output is a vector of K, each element represents the predicted probability of the corresponding category, and Σ_{j=1}^Kexp(z_j) represents the sum of the exponential function values of all elements in the input vector z.
