Signal Feature Perception Method Based on Multi-Dimensional Index Weighted Convolution

Abstract

This invention discloses a signal feature perception method based on multi-dimensional index weighted convolution. The method is implemented using a neural network, which is constructed based on a preprocessing layer, at least one feature extraction layer, and an adaptive pooling layer. The feature extraction layer includes multi-dimensional index weighted convolution. The preprocessing layer converts the binary file of the signal to be processed into an initial feature map. The multi-dimensional index weighted convolution weights the initial feature map based on the signal type of the signal to be processed, obtaining a weighted feature map. A normalization layer and an activation layer normalize and activate the weighted feature map, obtaining a normalized activated weighted feature map. The adaptive pooling layer pools and concatenates the normalized activated weighted feature map to obtain a global feature vector, and the signal feature perception result is determined based on the global feature vector. This invention, based on multi-dimensional index weighted convolution, improves the stability and generalization ability of feature extraction through the collaborative evaluation of multiple dimensions.

Description

Technical Field

[0001] This invention relates to the field of signal feature processing technology, and in particular to a signal feature perception method based on multi-dimensional index weighted convolution. Background Technology

[0002] Traditional convolutional neural networks have long relied on fixed-weight convolutional kernels for signal (e.g., acoustic signals) feature extraction. This static processing mechanism has fundamental technical flaws, including: 1. In terms of feature importance assessment, traditional convolutional layers use a uniform weight allocation strategy, ignoring the differences in importance of features across different channels. 2. Regarding dynamic adaptability, existing methods have shortcomings in signal processing: 1) Fixed-weight convolutional kernels struggle to adaptively adjust feature response intensity when facing changes in lighting conditions; 2) When the input is partially occluded, there is a lack of a mechanism to prioritize unoccluded features; 3) Limited adaptability to non-rigid deformations, making it impossible to dynamically adjust the weight distribution of the receptive field. 3. In terms of computational efficiency, traditional methods often require a significant increase in network depth and the number of parameters, leading to an exponential increase in computational complexity and reliance on complex attention modules and other additional structures, introducing additional computational overhead. Therefore, how to achieve efficient and accurate signal feature extraction has become an urgent problem to be solved.

[0003] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0004] The main objective of this invention is to provide a signal feature perception method based on multi-dimensional index weighted convolution, aiming to solve the technical problem of how to efficiently and accurately extract signal features.

[0005] To achieve the above objectives, the present invention provides a signal feature perception method based on multi-dimensional index weighted convolution. The signal feature perception method is implemented by a neural network, which is constructed based on a preprocessing layer, at least one feature extraction layer, and an adaptive pooling layer. The feature extraction layer is constructed by a combination of multi-dimensional index weighted convolution, a BatchNorm layer, and an activation layer.

[0006] The preprocessing layer converts the binary file of the signal to be processed into an initial feature map, wherein the signal to be processed is an acoustic signal;

[0007] The multi-dimensional index-weighted convolution performs weighted processing on the initial feature map based on the signal type of the signal to be processed, to obtain a weighted feature map;

[0008] The BatchNorm layer normalizes the weighted feature map to obtain a normalized weighted feature map.

[0009] The activation layer obtains a normalized activation weighted feature map based on the normalized weighted feature map using a preset activation function;

[0010] The adaptive pooling layer pools and concatenates the normalized activation-weighted feature map to obtain a global feature vector, and determines the signal feature perception result based on the global feature vector.

[0011] Optionally, the preprocessing layer is constructed by combining a Conv layer, a BatchNorm layer, and an activation layer;

[0012] The Conv layer performs standard convolution on the binary file of the signal to be processed to obtain an equal-dimensional mapping feature map;

[0013] The BatchNorm layer normalizes the equal-dimensional mapping feature map to obtain a normalized equal-dimensional feature map.

[0014] The activation layer obtains an initial feature map based on the normalized equal-dimensional feature map using the ReLU activation function.

[0015] Optionally, the step of weighting the initial feature map based on the signal type of the signal to be processed to obtain a weighted feature map includes:

[0016] Based on the signal type of the signal to be processed, the initial feature map is subjected to standard convolution processing to obtain a standard convolution feature map;

[0017] The weighted feature map is obtained by using the standard convolutional feature map and the weights of each channel through a weighting formula.

[0018] Optionally, before obtaining the weighted feature map based on the standard convolutional feature map and the channel weights using a weighting formula, the following steps are included:

[0019] Determine the batch size, channel index, feature map height, and feature map width for multi-dimensional index-weighted convolution;

[0020] Based on the batch size, the channel index, the feature map height, and the feature map width, construct the activation intensity index function, gradient index function, and noise sensitivity function for each channel, respectively.

[0021] A multi-dimensional index weighted constraint function is constructed using the sigmoid function based on the activation intensity index function, the gradient index function, and the noise sensitivity function.

[0022] The weights of each channel are calculated based on the multi-dimensional index weighted constraint function.

[0023] Optionally, the step of constructing the activation intensity index function, gradient index function, and noise sensitivity function for each channel based on the batch size, the channel index, the feature map height, and the feature map width includes:

[0024] Construct an activation intensity index function for each channel based on the batch size, the feature map height, and the feature map width;

[0025] The L2 norm of the standard convolutional feature map is obtained based on the channel index, the feature map height, and the feature map width. The gradient index function of each channel is constructed based on the L2 norm, the batch size, the feature map height, and the feature map width.

[0026] The standard convolutional feature map samples are subjected to noise perturbation to obtain a noise-perturbed feature map;

[0027] Based on the noise perturbation feature map and the standard convolutional feature map samples, a noise sensitivity function for each channel is constructed according to the batch size, the feature map height, and the feature map width.

[0028] Optionally, the adaptive pooling layer includes an AdaptiveMaxPool2d layer, an AdaptiveAvgPool2d layer, and a splicing layer;

[0029] The AdaptiveMaxPool2d layer performs adaptive max pooling on the normalized activation-weighted feature map to obtain a max pooled feature vector.

[0030] The AdaptiveAvgPool2d performs an adaptive average pooling operation on the normalized activation-weighted feature map to obtain an average pooled feature vector.

[0031] The concatenation layer concatenates the max-pooling feature vector and the average-pooling feature vector based on the channel dimension to obtain a global feature vector.

[0032] Furthermore, to achieve the above objectives, the present invention also proposes a storage medium storing a signal recognition program, which, when executed by a processor, implements the steps of the signal feature perception method based on multi-dimensional index weighted convolution as described above.

[0033] This invention discloses a signal feature perception method based on multi-dimensional index weighted convolution. The method is implemented using a neural network, which is constructed from a preprocessing layer, at least one feature extraction layer, and an adaptive pooling layer. The feature extraction layer includes multi-dimensional index weighted convolution. First, the preprocessing layer converts the binary file of the signal to be processed (the signal is acoustic). Then, the multi-dimensional index weighted convolution weights the initial feature map based on the signal type of the signal to be processed, obtaining a weighted feature map. Next, a BatchNorm layer normalizes the weighted feature map, resulting in a normalized weighted feature map. An activation layer then uses a preset activation function to obtain a normalized activation weighted feature map. Finally, an adaptive pooling layer pools and concatenates the normalized activation weighted feature maps to obtain a global feature vector. The signal feature perception result is determined based on this global feature vector. This invention, using a neural network constructed based on multi-dimensional index weighted convolution, improves the stability and generalization ability of feature extraction through the collaborative evaluation of multiple dimensions. Attached Figure Description

[0034] Figure 1 This is a structural block diagram of the first embodiment of the signal feature perception method based on multi-dimensional index weighted convolution of the present invention;

[0035] Figure 2 This is a structural block diagram of multiple feature extraction layers in the first embodiment of the signal feature perception method based on multi-dimensional index weighted convolution of the present invention;

[0036] Figure 3 This is a structural block diagram of the signal recognition model of the first embodiment of the signal feature perception method based on multi-dimensional index weighted convolution of the present invention;

[0037] Figure 4 This is a schematic diagram of weight constraints in the first embodiment of the signal feature perception method based on multi-dimensional index weighted convolution of the present invention.

[0038] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0039] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0040] This invention provides a signal feature perception method based on multi-dimensional index-weighted convolution, referring to... Figure 1 , Figure 1 This is a structural block diagram of the first embodiment of the signal feature perception method based on multi-dimensional index weighted convolution of the present invention.

[0041] In this embodiment, the signal feature perception method is implemented by a neural network. The neural network is constructed based on a preprocessing layer 101, at least one feature extraction layer 102, and an adaptive pooling layer 103. The feature extraction layer is constructed by a combination of a multi-dimensional metric weighted (MDMW) convolution 1021, a batchNorm layer 1022, and an activation layer 1023.

[0042] The preprocessing layer is constructed by combining a Conv layer, a BatchNorm layer, and an activation layer. The Conv layer is implemented using standard PyTorch convolutions, with the number of input channels equal to the number of output channels. The corresponding 1D or 2D PyTorch standard convolution is selected based on the signal type. At the same time, the corresponding 1D or 2D BatchNorm layer is determined to perform normalization. The activation function in the activation layer is the ReLU function.

[0043] It should be understood that the signal can be an acoustic signal, an optical signal, an electrical signal, etc. An acoustic signal can be an underwater acoustic signal, etc.

[0044] The signal type can be 1D or 2D. 1D signals can be time-series data such as audio, while 2D signals can be two-dimensional data such as images (the images are grayscale images).

[0045] The feature extraction layer is constructed from a combination of multi-dimensional index-weighted convolutions, BatchNorm layers (i.e., normalization layers), and activation layers. In the specific implementation, the corresponding multi-dimensional index-weighted convolutions can be selected according to the signal type, and the corresponding 1D or 2D BatchNorm layers can be determined to perform normalization. The activation function in the activation layer can be either the ReLU function or the GELU function.

[0046] It should also be noted that the multi-dimensional index weighted convolution is a dynamic convolution constructed based on a combination of multi-dimensional index weighting constraint function and PyTorch standard convolution (i.e., Conv layer).

[0047] In this embodiment, a single feature extraction layer can be set within the neural network. The activation layer within the feature extraction layer uses the ReLU function, as referenced. Figure 2 , Figure 2This is a structural block diagram of multiple feature extraction layers in the first embodiment of the signal feature perception method based on multi-dimensional index weighted convolution of the present invention. Multiple feature extraction layers can also be set in the neural network. If the multiple feature extraction layers are feature extraction layer A, feature extraction layer B, ... feature extraction layer N, the preprocessing layer inputs the feature map into feature extraction layer A for processing to obtain a normalized activation-weighted feature map. Then, the normalized activation-weighted feature map is input into feature extraction layer B for processing again until feature extraction layer N outputs the final normalized activation-weighted feature map. The normalized activation-weighted feature map output by feature extraction layer N is then input into the adaptive pooling layer for subsequent processing.

[0048] It should also be noted that the reference Figure 2 It is recommended to use the ReLU function in the shallow layers (i.e., the activation layers corresponding to feature extraction layers A...M) and the GELU function in the deep layers (i.e., the activation layers corresponding to feature extraction layers L...N). The activation function layer in the intermediate feature extraction layers can be selected as needed.

[0049] It should also be noted that the feature extraction layers A...M are the user-defined preset number of feature extraction layers (from the beginning A to M), and the feature extraction layers L...N are the user-defined preset number of feature extraction layers (from L to the end N).

[0050] There may be activation layers corresponding to other feature extraction layers between feature extraction layers M and L. If there are no activation layers corresponding to other feature extraction layers, then feature extraction layers M and L are the same feature extraction layer.

[0051] The adaptive pooling layer includes the AdaptiveMaxPool2d layer, the AdaptiveAvgPool2d layer, and the splicing layer.

[0052] In this embodiment, reference Figure 3 , Figure 3 This is a structural block diagram of the signal recognition model in the first embodiment of the signal feature perception method based on multi-dimensional index weighted convolution of the present invention. The classification head can be used to optimize the network parameters within the neural network, and the neural network and the classification head can also be combined to construct a signal recognition model. The classification head is constructed by combining a Flatten layer, a fully connected layer, a BatchNorm layer, an activation layer, a Dropout layer, a fully connected layer, and a Softmax layer: Flatten layer + fully connected layer + BatchNorm layer + activation layer + Dropout layer + fully connected layer + Softmax layer.

[0053] In this embodiment, the Flatten layer flattens the global feature vector into a two-dimensional feature vector with batch dimension. The fully connected layer reduces the feature dimension of the two-dimensional feature vector to obtain the feature vector. The BatchNorm layer normalizes the feature vector obtained after dimensionality reduction to obtain a normalized feature vector. The activation layer activates the normalized feature vector through GELU to obtain an activated feature vector. The Dropout layer randomly drops the activated feature vector. The fully connected layer reduces the dimension of the vector after random dropping to the number of categories (e.g., N=4) and outputs the feature shape (N). The Softmax layer can choose Sigmoid for 2-class classification and obtains the final signal recognition result (e.g., cargo ship).

[0054] In the specific implementation, the training process is as follows: the binary data samples corresponding to the signal samples are input into the initial neural network for training. The preprocessing layer converts the binary data samples into initial feature map samples. The initial feature map samples are then subjected to PyTorch standard convolution by multi-dimensional index weighted convolution of the signal type corresponding to the signal sample, resulting in standard convolution feature map samples. The standard convolution feature map samples include multi-channel feature map samples, and the feature map of its c-th channel is y[b,c,h,w]=conv. c (x), then determine the batch size, channel index, feature map height, and feature map width corresponding to the multi-dimensional index weighted convolution; construct the activation intensity index function, gradient index function, and noise sensitivity function for each channel according to the batch size, channel index, feature map height, and feature map width respectively; construct the multi-dimensional index weighted constraint function (i.e., W[c]=sigmoid(F[c])+b; F[c]=αA[c]+βG[c]+γS[c], where b, α, β, and γ are the parameters corresponding to the weights, which can be customized by the user. For example, b=0.4, α=0.5, β=0.3, γ=0.2); calculate the weight W[c] of each channel based on the multi-dimensional index weighted constraint function, and then calculate the weighted feature map sample P (i.e., P=) based on the weight of each channel and the feature image samples of each channel in the channel feature map sample. The BatchNorm layer normalizes the weighted feature map samples to obtain normalized weighted feature map samples. The activation layer obtains normalized activated weighted feature map samples by using a preset activation function (RELU or GELU) based on the normalized weighted feature map samples. The adaptive pooling layer pools and concatenates the normalized activated weighted feature map samples to obtain global feature vector samples. Finally, the global feature vector samples are input to the classification head, and the classification head outputs the signal recognition result samples corresponding to the signal samples. Based on the signal recognition result samples and the corresponding real recognition results, the loss function value corresponding to the initial recognition model is calculated using a loss function. Based on the loss function value, the model parameters (i.e., the weights of each channel) in the initial neural network are updated in reverse until the loss function value is less than or equal to a preset threshold.

[0055] It should also be noted that W will be updated in real time during training as the following three metrics change, while the channel weights of the trained neural network will not change. (Reference) Figure 4 , Figure 4 This is a schematic diagram of weight constraints in the first embodiment of the signal feature perception method based on multi-dimensional index weighted convolution of the present invention. After training, the channel weights of the multi-dimensional index weighted convolution in the neural network can be directly applied.

[0056] Furthermore, the processing method for constructing the activation intensity index function, gradient index function, and noise sensitivity function for each channel based on batch size, channel index, feature map height, and feature map width is as follows: The activation intensity index function for each channel is constructed based on batch size, feature map height, and feature map width; the L2 norm of the standard convolutional feature map is obtained based on the channel index, feature map height, and feature map width, and the gradient index function for each channel is constructed based on the L2 norm, the batch size, feature map height, and feature map width; noise perturbation is applied to the standard convolutional feature map samples to obtain a noise perturbation feature map; and the noise sensitivity function for each channel is constructed based on the noise perturbation feature map and the standard convolutional feature map samples, according to the batch size, feature map height, and feature map width.

[0057] Activation intensity index function:

[0058]

[0059] In the formula, Let y[b,c,h,w] be the activation intensity index value of channel c. c(x) represents the output of the c-th channel of the standard convolution, h is the height of the feature map sample corresponding to the c-th channel, b is the batch size of the feature map sample corresponding to the c-th channel, w is the width of the feature map sample corresponding to the c-th channel, B is the total batch size, C is the total channel index, H is the total height of the feature map, and W is the total width of the feature map.

[0060] Gradient index function:

[0061]

[0062]

[0063] In the formula, Let |y|2 be the gradient index value of a certain channel, and let |y|2 be the L2 norm of the standard convolution feature map sample output by the standard convolution.

[0064] Noise sensitivity function:

[0065]

[0066] In the formula, The noise sensitivity of a certain channel. To add noise perturbation to the standard convolution output (i.e., the noise perturbation feature map), the noise... (N(0,1) is a standard normal distribution).

[0067] It should also be noted that, similar to A[c], G[c], This way of describing it refers to the calculation of the index of a single channel. It is the same function for all channels in the entire MDMWConv layer. After the calculation of each channel, they are concatenated into a weight matrix, which is passed in the network in the form of a PyTorch tensor and participates in the calculation of the final output feature map.

[0068] In practical applications, the processing method for signal feature sensing is as follows:

[0069] The preprocessing layer converts the binary file of the signal to be processed into an initial feature map.

[0070] It should be understood that the signal to be processed can be an acoustic signal, an optical signal, an electrical signal, etc. The signal type corresponding to the signal to be processed is determined in advance. The signal type can be 1D or 2D. 1D signals correspond to time-series data such as audio, while 2D signals correspond to two-dimensional data such as images (images are grayscale images). Then, the signal to be processed undergoes data preprocessing (for example, if the signal corresponds to 1D, the sampling rate is standardized; if it corresponds to 2D, the resolution is standardized), and the preprocessed data is converted into a binary file. This binary file is then input to the preprocessing layer, which converts the binary file of the signal to be processed into an initial feature map.

[0071] It should be noted that the preprocessing layer is constructed by combining the Conv layer, the BatchNorm layer, and the activation layer; the Conv layer performs standard convolution on the binary file of the signal to be processed to obtain an equal-dimensional mapping feature map; the BatchNorm layer normalizes the equal-dimensional mapping feature map to obtain a normalized equal-dimensional feature map.

[0072] The activation layer obtains the initial feature map based on the normalized equal-dimensional feature map through the ReLU activation function.

[0073] The initial feature map is sent to the feature extraction layer, which is constructed by combining multi-dimensional index-weighted convolution, BatchNorm layer and activation layer.

[0074] Multi-dimensional index-weighted convolution performs standard convolution processing on the initial feature map based on the signal type of the signal to be processed, resulting in a standard convolution feature map. A weighted feature map is then obtained using a weighting formula based on the standard convolution feature map and the weights of each channel. The BatchNorm layer normalizes the weighted feature map, resulting in a normalized weighted feature map. The activation layer then uses a preset activation function to obtain a normalized activation-weighted feature map based on the normalized weighted feature map.

[0075] The normalized activation-weighted feature map is sent to the adaptive pooling layer, which includes the AdaptiveMaxPool2d layer, the AdaptiveAvgPool2d layer, and the connection layer.

[0076] The AdaptiveMaxPool2d layer performs adaptive max pooling on the normalized activation-weighted feature map to obtain the max pooled feature vector; the AdaptiveAvgPool2d layer performs adaptive average pooling on the normalized activation-weighted feature map to obtain the average pooled feature vector; the connection layer concatenates the max pooled feature vector and the average pooled feature vector based on the channel dimension to obtain the global feature vector.

[0077] It should also be understood that users can choose the corresponding connection layer according to the architecture they are using. For example, torch uses the cat function to add tensors with the same dimensions in other dimensions in the channel dimension (e.g., (1, 1, 64) and (1, 1, 64) are concatenated to (1, 2, 64)).

[0078] It should be noted that the global feature vector can be used as the result of signal feature perception. The global feature vector can be input into the classification head for signal recognition, and can also be used for other subsequent data processing.

[0079] In this embodiment, the signal feature perception method is implemented by a neural network. The neural network is constructed based on a preprocessing layer, at least one feature extraction layer, and an adaptive pooling layer. The feature extraction layer includes a multi-dimensional index-weighted convolution. First, the preprocessing layer converts the binary file of the signal to be processed into an initial feature map. Then, the multi-dimensional index-weighted convolution weights the initial feature map based on the signal type of the signal to be processed to obtain a weighted feature map. Next, the BatchNorm layer normalizes the weighted feature map to obtain a normalized weighted feature map. The activation layer obtains a normalized activation-weighted feature map based on the normalized weighted feature map through a preset activation function. Finally, the adaptive pooling layer pools and concatenates the normalized activation-weighted feature maps to obtain a global feature vector, and the signal feature perception result is determined based on the global feature vector. This embodiment, based on the neural network constructed by the multi-dimensional index-weighted convolution, improves the stability and generalization ability of feature extraction through the collaborative evaluation of multi-dimensional indices.

[0080] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0081] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0082] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory / random access memory, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0083] The above are merely preferred embodiments of the present invention and do not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A signal feature perception method based on multi-dimensional index-weighted convolution, characterized in that, The signal feature perception method is implemented by a neural network, which is constructed based on a combination of a preprocessing layer, at least one feature extraction layer and an adaptive pooling layer. The feature extraction layer is constructed by a combination of multi-dimensional index weighted convolution, a BatchNorm layer and an activation layer. The multi-dimensional index weighted convolution is a dynamic convolution constructed based on a combination of a multi-dimensional index weighted constraint function and a Conv layer. The preprocessing layer converts the binary file of the signal to be processed into an initial feature map; The multi-dimensional index-weighted convolution performs weighted processing on the initial feature map based on the signal type of the signal to be processed, to obtain a weighted feature map; The BatchNorm layer normalizes the weighted feature map to obtain a normalized weighted feature map. The activation layer obtains a normalized activation weighted feature map based on the normalized weighted feature map using a preset activation function; The adaptive pooling layer pools and concatenates the normalized activation-weighted feature map to obtain a global feature vector, and determines the signal feature perception result based on the global feature vector. The step of weighting the initial feature map based on the signal type of the signal to be processed to obtain a weighted feature map includes: Based on the signal type of the signal to be processed, the initial feature map is subjected to standard convolution processing to obtain a standard convolution feature map; Determine the batch size, channel index, feature map height, and feature map width for multi-dimensional index-weighted convolution; Based on the batch size, the channel index, the feature map height, and the feature map width, construct the activation intensity index function, gradient index function, and noise sensitivity function for each channel, respectively. A multi-dimensional index weighted constraint function is constructed using the sigmoid function based on the activation intensity index function, the gradient index function, and the noise sensitivity function. The weights of each channel are calculated based on the multi-dimensional index weighting constraint function. The weighted feature map is obtained by using the standard convolutional feature map and the weights of each channel through a weighting formula.

2. The method as described in claim 1, characterized in that, The preprocessing layer is constructed by combining a Conv layer, a BatchNorm layer, and an activation layer. The Conv layer performs standard convolution on the binary file of the signal to be processed to obtain an equal-dimensional mapping feature map; The BatchNorm layer normalizes the equal-dimensional mapping feature map to obtain a normalized equal-dimensional feature map. The activation layer obtains an initial feature map based on the normalized equal-dimensional feature map using the ReLU activation function.

3. The method as described in claim 1, characterized in that, The step of constructing the activation intensity index function, gradient index function, and noise sensitivity function for each channel based on the batch size, the channel index, the feature map height, and the feature map width includes: Construct an activation intensity index function for each channel based on the batch size, the feature map height, and the feature map width; The L2 norm of the standard convolutional feature map is obtained based on the channel index, the feature map height, and the feature map width. The gradient index function of each channel is constructed based on the L2 norm, the batch size, the feature map height, and the feature map width. The standard convolutional feature map samples are subjected to noise perturbation to obtain a noise-perturbed feature map; Based on the noise perturbation feature map and the standard convolutional feature map samples, a noise sensitivity function for each channel is constructed according to the batch size, the feature map height, and the feature map width.

4. The method as described in claim 1, characterized in that, The adaptive pooling layer includes an AdaptiveMaxPool2d layer, an AdaptiveAvgPool2d layer, and a splicing layer; The AdaptiveMaxPool2d layer performs adaptive max pooling on the normalized activation-weighted feature map to obtain a max pooled feature vector. The AdaptiveAvgPool2d performs an adaptive average pooling operation on the normalized activation-weighted feature map to obtain an average pooled feature vector. The concatenation layer concatenates the max-pooling feature vector and the average-pooling feature vector based on the channel dimension to obtain a global feature vector.

Images