Target identification method and device based on low-rank neural network

Abstract

The invention provides a target recognition method and device based on a low-rank neural network, and relates to the field of radar target recognition. According to the scheme, CP decomposition is carried out on a convolution kernel weight tensor of a two-dimensional convolution kernel in a convolution module to obtain a low-rank convolution module; the decomposition rank of CP decomposition is determined according to the Kruskal uniqueness theory and the computing power of low-rank neural network operation equipment; performing matrix decomposition on the full-connection layer weight matrix of the full-connection module to obtain a low-rank full-connection module; the decomposition rank of matrix decomposition is determined according to the input feature dimension and the output feature dimension of the full-connection module; replacing a convolution module and a full connection module in the convolutional neural network with a low-rank convolution module and a low-rank full connection module to obtain a low-rank neural network; and collecting target feature parameters, inputting the target feature parameters into the low-rank neural network, and obtaining a target recognition result. By using the method, relatively small performance loss can be kept under the condition that the model parameter quantity is greatly compressed.

Description

Technical Field

[0001] The present invention relates to the field of radar target recognition, and particularly to a target recognition method and device based on a low-rank neural network. Background Art

[0002] Radar target recognition refers to a technology that irradiates a target with electromagnetic waves emitted by a radar and analyzes the received echo to determine the type of the target. The radar data commonly used for target recognition mainly includes radar data with good classification features such as Synthetic Aperture Radar (SAR) images, Inverse synthetic aperture radar (ISAR) images, one-dimensional Radar Cross section (RCS) data and its time-frequency transformation diagrams, and one-dimensional range profile data. In scenarios where quick decision-making is required, rapid recognition of radar targets is demanded. The spaceborne radar target recognition algorithm can achieve quick response and rapid recognition, and at the same time only needs to transmit a small amount of algorithm result data to the ground radar station instead of a large amount of raw data, which is beneficial for quick decision-making on the ground.

[0003] The currently widely used Convolutional Neural Network (CNN) models are usually storage-intensive and computation-intensive, requiring huge storage resources and computational costs. However, the resources of spaceborne computing platforms are limited and they cannot deploy large target recognition models, making it difficult to effectively deploy them on spaceborne computing platforms. To solve this problem, researchers have proposed many model compression methods, and low-rank decomposition is a commonly used compression technique among them.

[0004] Due to the particularity of the spaceborne target recognition task, not only is there a strict requirement for the size of the radar target recognition model, but also the accuracy of target recognition needs to be ensured. Most of the traditional neural network model compression algorithms based on low-rank decomposition decompose the parameters of the pre-trained model, decompose the high-dimensional tensor into a combination of multiple low-rank factor matrices, and use this low-rank property to approximate the original weights. The traditional low-rank decomposition algorithm requires multiple iterative fine-tuning, has obvious convergence problems, high training costs, and the selection of the decomposition rank mostly starts from manual experience, lacking certain constraints. The performance loss when the model compression ratio is high cannot meet the needs of the spaceborne radar target recognition task. Summary of the Invention

[0005] In view of this, the present invention provides a target recognition method and device based on a low-rank neural network, which can maintain a small performance loss while greatly compressing the number of model parameters.

[0006] To solve the above technical problems, the present invention is implemented as follows.

[0007] A target recognition method based on a low-rank neural network, comprising: Step 1: Design a low-rank neural network: Design a low-rank convolution module: perform CP decomposition on the convolution kernel weight tensor of the two-dimensional convolution kernel in the convolution module; determine the decomposition rank of the convolution kernel weight tensor in the CP decomposition according to the Kruskal uniqueness theory and the computing power of the low-rank neural network operating device; obtain the low-rank convolution module through CP decomposition; Design a low-rank fully connected module: perform matrix decomposition on the weight matrix of the fully connected layer of the fully connected module; determine the decomposition rank of the weight matrix of the fully connected layer according to the input feature dimension and output feature dimension of the fully connected module; obtain the low-rank fully connected module through matrix decomposition; Adopt the low-rank convolution module and the low-rank fully connected module to replace the convolution module and the fully connected module in the convolutional neural network to obtain a low-rank neural network; Step 2: Train the low-rank neural network: train the low-rank neural network using training samples without pre-training the network weights; Step 3: Collect target feature parameters and input them into the trained low-rank neural network, and the low-rank neural network outputs the target recognition result.

[0008] Preferably, the CP decomposition of the convolution kernel weight tensor of the two-dimensional convolution kernel in the convolution module is as follows: CP decompose the convolution kernel weight tensor into factor matrices , , , , where , , , are the rank-1 vectors constituting the factor matrices , , , respectively, , is the decomposition rank of the convolution kernel weight tensor; The four dimensions of the two-dimensional convolution kernel are the output channel dimension, the input channel dimension, the convolution kernel height, and the convolution kernel width, and the convolution kernel height and the convolution kernel width are the same; The convolution kernel weight tensor is denoted as , and performing CP decomposition on gives:

[0009] where, " " represents the vector outer product operation; is an element in the convolutional kernel weight tensor , are respectively elements in the vector ; The subscript o represents the o-th in the number of output channels of the two-dimensional convolutional kernel, and the subscript i represents the i -th in the number of input channels of the two-dimensional convolutional kernel, and the subscript k is the k -th in the convolutional kernel height or convolutional kernel width dimension; The matrix factorization of the weight matrix of the fully connected layer of the fully connected module is as follows: The weight matrix of the fully connected layer is respectively the factor matrix U and the factor matrix V; Let be the weight matrix of the fully connected layer. Matrix factorization of gives:

[0010] wherein, is an element in the weight matrix of the fully connected layer , are respectively elements in the factor matrix ; The subscript out represents the out -th in the output feature dimension of the fully connected module, and the subscript in represents the in -th in the input feature dimension of the fully connected module; The subscript r represents the -th rank-1 vector in the factor matrix r , is the decomposition rank of the weight matrix of the fully connected layer.

[0011] Preferably, the decomposition rank of the convolutional kernel weight tensor in the CP decomposition is determined according to the Kruskal uniqueness theory and the computing power of the low-rank neural network running device as follows: Determine the first decomposition rank according to the Kruskal uniqueness theory; Determine the maximum rank according to the computing power of the low-rank neural network running device; The lower the computing power of the low-rank neural network running device, the higher the compression ratio required, the smaller the selection; The higher the computing power of the low-rank neural network running device, the larger the selection; Take the minimum value of the first decomposition rank and the maximum rank as the decomposition rank of the convolutional kernel weight tensor.

[0012] Preferably, the first decomposition rank is determined according to the Kruskal uniqueness theory as follows: The convolution kernel weight tensor is a fourth-order tensor, and the first decomposition rank is deduced according to the Kruskal uniqueness theory satisfying the following conditions:

[0013] wherein, is the number of output channels of the two-dimensional convolution kernel; is the number of input channels of the two-dimensional convolution kernel; is the height and width of the convolution kernel in the two-dimensional convolution kernel, and the height is the same as the width.

[0014] Preferably, according to the input feature dimension and the output feature dimension of the fully connected module, the decomposition rank of the weight matrix of the fully connected layer is: the decomposition rank of the weight matrix of the fully connected layer takes the output feature dimension of the fully connected module.

[0015] Preferably, in step 2, during the training process, the network parameters of the low-rank neural network are initialized by Kaiming; starting from the initialized network parameters, the low-rank neural network is trained without pre-training the network weights.

[0016] Preferably, the low-rank neural network includes a preliminary feature extraction module, a multi-feature fusion module, and a feature re-integration and classification module; The preliminary feature extraction module is used to perform preliminary feature extraction on the input target feature parameters, and the extracted preliminary features are input into the multi-feature fusion module; The multi-feature fusion module is composed of N fusion extraction networks with the same structure connected in series; each fusion extraction network includes a shallow feature extraction channel, a deep feature extraction channel, an attention feature extraction channel, and a splicing fusion unit; the shallow feature extraction channel has feature extraction layers, and the preliminary features are output as shallow features after passing through the shallow feature extraction channel; the deep feature extraction channel has feature extraction layers, , and the preliminary features are output as deep features after passing through the deep feature extraction channel; In the attention feature extraction channel: the first branch performs convolution feature extraction and expansion on the preliminary features to obtain the first branch features ; the second branch performs convolution feature extraction and expansion on the preliminary features , and then is processed by a Sigmoid function layer to obtain the parameter amplitude features ; The layer-decreasing mask module D-Mask performs a partial zeroing operation on the deep features and outputs a feature mask ; As the layers of the fusion extraction network gradually deepen from 1 to N , the representativeness of the features is continuously enhanced, and the layer-decreasing mask module gradually attenuates the zeroing ratio of the deep features ; The first branch feature , the parameter amplitude feature and the feature mask are multiplied to generate an attention feature ; The layer-decreasing mask module D-Mask is only used during network training; The shallow features , the deep features and the attention feature are feature-fused to output a fused feature , which is input to the feature re-integration and classification module; The feature re-integration and classification module processes the fused feature and outputs the target recognition result.

[0017] Preferably, the way that the layer-decreasing mask module D-Mask performs a partial zeroing operation on the deep features and outputs a feature mask is as follows: For N the th layer in the layer fusion extraction network, calculate the zeroing ratio

[0018] where and are respectively the zeroing ratios of the set first layer and the last layer of the fusion extraction network; According to the zeroing ratio , perform a partial zeroing judgment on each element in the deep features output from the deep feature extraction channel of the th layer of the fusion extraction network: Judge whether the absolute value of the element in the deep features is greater than or equal to the absolute value threshold set for the th layer; and generate a uniformly distributed random variable, and judge whether the uniformly distributed random variable is less than ; If both the judgment of the element absolute value and the uniformly distributed random variable are yes, set the position of the current element in the feature mask to 0, otherwise, set it to 1; By performing the partial zeroing judgment operation on all elements of the deep features , generate a feature mask 。

[0019] Preferably, when training the low-rank neural network, construct an isomorphic convolutional neural network without replacing the low-rank convolutional module and the low-rank fully-connected module. First, train the isomorphic convolutional neural network as a teacher model, and then form a distillation learning network with the trained teacher model and the low-rank neural network to perform distillation learning on the low-rank neural network.

[0020] The present invention also provides an object recognition device based on a low-rank neural network, which includes a low-rank neural network and a training module; The low-rank neural network is obtained by replacing the convolutional module in the convolutional neural network with a low-rank convolutional module and the fully-connected module with a low-rank fully-connected module; the collected target feature parameters are input into the low-rank neural network, and the low-rank neural network outputs the target recognition result; The determination method of the low-rank convolutional module is as follows: perform CP decomposition on the convolutional kernel weight tensor of the two-dimensional convolutional kernel in the convolutional module; determine the decomposition rank of the convolutional kernel weight tensor in the CP decomposition according to the Kruskal uniqueness theory and the computing power of the low-rank neural network operation device; obtain the low-rank convolutional module through CP decomposition; The determination method of the low-rank fully-connected module is as follows: perform matrix decomposition on the weight matrix of the fully-connected layer of the fully-connected module; determine the decomposition rank of the weight matrix of the fully-connected layer according to the input feature dimension and the output feature dimension of the fully-connected module; obtain the low-rank fully-connected module through matrix decomposition; The training module is used to train the low-rank neural network with training samples without pre-training the network weights.

[0021] Beneficial effects: (1) The present invention proposes a lightweight neural network design scheme CLRND based on CP decomposition and matrix decomposition. Different from most methods that only perform low-rank decomposition on a certain block of parameters, the present invention designs low-rank convolutional modules and low-rank fully-connected modules according to different decomposition methods by using the differences in the forms and dimensions of parameters in different layers, and can maintain a small performance loss while greatly compressing the model parameters.

[0022] Secondly, for the selection of the decomposition rank in traditional other algorithms, especially for the low-rank decomposition of high-order tensors, it mostly starts from manual experience. The present invention starts from the uniqueness theory and actual compression requirements in the rank decomposition of high-order tensors, rationalizes the selection of the rank, and also derives the selection of the matrix decomposition rank mathematically, and combines the actual data characteristics to select a reasonable rank.

[0023] In addition, different from the prior art which requires high-rank information of a pre-trained model, the present invention only needs to initialize the network parameters, allowing the neural network to learn the effective information of the low-rank subspace from scratch, making the low-rank approximation more flexible, with rapid convergence and less computational overhead.

[0024] (2) The low-rank convolutional module LRCPC proposed by the present invention performs low-rank decomposition on the convolutional kernel in the form of a high-order tensor, which can better utilize the structural information of the high-order tensor and does not require dimension reshaping operations on the weight tensor, reducing intermediate operation steps; at the same time, starting from the theory of uniqueness of high-order tensor rank decomposition and combining with the actual compression capacity requirements, the decomposition rank is constrained within a reasonable range, preferably avoiding the problems of blind selection of the decomposition rank and poor interpretability.

[0025] (3) The low-rank fully connected module LRCPFc proposed by the present invention simplifies the decomposition by getting rid of the complex decomposition process of the weight matrix of the pre-trained model according to the characteristics of the weight matrix of the fully connected layer, derives the rank selected for this decomposition mathematically, and reasonably limits the size of the decomposition rank according to the actual application needs and general data characteristics. This enables the network to learn effective information under this condition while reducing the stacking of redundant information.

[0026] (4) In a preferred embodiment, an attention feature extraction channel implemented by a filter balance attention module FBAM is proposed to enhance the perception ability of key features, enabling the network to focus on key features; and a large weight parameter of the layer-by-layer decreasing mask module D-mask is randomly set to zero, and the random zeroing ratio is different for different layers of the network. As the network layer deepens, the feature representativeness is continuously strengthened, and the zeroing ratio decays layer by layer, thereby retaining more high-semantic features, making the model more flexible in mining feature information, avoiding the fitting problem caused by parameter homogenization, and finally enabling the network to learn more potential and robust features.

[0027] (5) As the number of layers of the traditional CNN model increases, feature representation deviation often occurs, and the model effect even degrades. To solve this problem, the inventors of the present invention adopt a multi-feature fusion scheme in the low-rank neural network, fusing the shallow features, deep features, and attention features of the model to enhance the feature expression ability, effectively avoiding the problem of biased feature extraction of a single branch, and avoiding gradient disappearance and gradient explosion, thereby improving the training stability and robustness of the network.

[0028] (6) In a preferred embodiment, to address the problem that the speed of learning effective information representation from scratch is slow, a low-rank filter balance attention distillation network LFBADN is designed, which fuses multiple features to enhance the feature expression ability, and is guided by a teacher model with a larger number of parameters to quickly learn effective parameters for the low-rank neural network, improving the recognition rate. Brief Description of the Drawings

[0029] Figure 1 This is a schematic diagram of the principle of the object recognition method based on the low-rank neural network of the present invention.

[0030] Figure 2 This is a schematic diagram of the low-rank convolution module in the first embodiment of the present invention.

[0031] Figure 3 This is a schematic diagram of the forward process of the low-rank convolution module.

[0032] Figure 4 This is a schematic diagram of the low-rank fully connected module in the first embodiment of the present invention.

[0033] Figure 5 This is a schematic diagram of the forward process of the low-rank fully connected module.

[0034] Figure 6 This is a block diagram of the composition of the low-rank neural network in the second embodiment of the present invention.

[0035] Figure 7 For Figure 6 An example of the low-rank neural network shown.

[0036] Figure 8 This is a structural diagram of the filter balanced attention distillation network in the third embodiment of the present invention.

[0037] Figure 9 This is a block diagram of the composition of the object recognition device based on the low-rank neural network in the fourth embodiment of the present invention. Detailed implementation manners

[0038] The present invention will be described in detail below with reference to the accompanying drawings and by way of examples.

[0039] The present invention provides an object recognition method based on a low-rank neural network. This method adopts a lightweight neural network design method (Concise Low-Rank Network Design, CLRND) based on CP (CANDECOMP / PARAFAC) decomposition and matrix decomposition, abandons the relevant information in the full-rank high-precision parameters of the pre-trained model, and directly designs a low-rank neural network, including a low-rank convolution module (Low-rank CP Convolution, LRCPC) and a low-rank fully connected module (Low-rank CP Fully Connected Layer, LRCPFc). It enables the network to adaptively learn the factor matrix under the constraints of the optimal rank constraint and the computational complexity constraint, and trains the low-rank model from scratch to achieve better compression performance and model performance.

[0040] Embodiment 1 This embodiment provides an object recognition method based on a low-rank neural network, as Figure 1As shown in the figure, it includes the following steps: Step 1: Design a low-rank neural network.

[0041] The existing technology is based on the CP decomposition of the high-dimensional weight tensor of the pre-trained model, which has convergence problems, high training costs, and serious performance degradation when the compression ratio is high. Secondly, the existing technology does not have a good finite algorithm for determining the tensor rank. The classical Alternating Least Squares (ALS) algorithm cannot guarantee convergence to the global minimum, and due to the existence of degenerate tensors, it is difficult to find the optimal rank approximation. Moreover, the method based on hierarchical iteration and fine-tuning is also difficult to achieve balance in information utilization between layers. In addition, most low-rank decompositions based on pre-trained models use Singular Value Decomposition (SVD) operations or Bayesian Estimates, with a large computational overhead. Moreover, most low-rank decompositions are only for a certain type of network layer.

[0042] The present invention designs a low-rank neural network, which performs different low-rank decompositions for different types of modules. Among them, the CP decomposition is used for the convolutional module, and the convolutional kernel is decomposed in a low rank in the form of a high-order tensor, which can better utilize the structural information of the high-order tensor and does not require dimension reshaping operations on the weight tensor, reducing intermediate calculation steps. At the same time, starting from the uniqueness theory of high-order tensor rank decomposition and combining the actual compression capacity requirements, the decomposition rank is constrained within a reasonable range, which better avoids the problems of blind selection of the decomposition rank and poor interpretability. For the fully connected module, matrix decomposition is used. This design is based on the characteristics of the weight matrix of the fully connected layer, getting rid of the complex decomposition process of the weight matrix of the pre-trained model, simplifying the decomposition, deriving the rank selected for this decomposition mathematically, and reasonably restricting the size of the decomposition rank according to the actual application needs and general data characteristics, so that the network can learn effective information under this condition while reducing the stacking of redundant information.

[0043] The following will describe in detail the low-rank decomposition scheme and decomposition rank selection for the low-rank convolutional module and the low-rank fully connected module.

[0044] (1) Design a low-rank convolutional module (LRCPC): The four dimensions of the two-dimensional convolutional kernel in the convolutional module are: the output channel dimension , the input channel dimension , the convolutional kernel height , and the convolutional kernel width . According to the habit, the height and width of the convolutional kernel are set to the same value, that is = = K .

[0045] In the past, the decomposition of convolutional kernel parameters mostly involved matrix decomposition operations after matrixizing them. The present invention proposes a low-rank convolutional module that performs low-rank decomposition on convolutional kernel parameters in the form of high-order tensors to generate a low-rank convolutional structure, which can better utilize the structural information of high-dimensional tensors and does not require dimensional reconstruction (reshape) operations on the weight parameter tensors, reducing intermediate calculation steps.

[0046] As Figure 2 shown, the CP decomposition of the convolutional kernel weight tensor in the convolutional module is as follows: The CP decomposition of the convolutional kernel weight tensor into factor matrices , , , , where , , , are the rank-1 vectors that make up the factor matrices , , , , respectively; , , , ; , is the decomposition rank of the convolutional kernel weight tensor; , .

[0047] is the convolutional kernel weight tensor, , for performing CP decomposition gives: (1) where, " " represents the vector outer product operation; is an element in the convolutional kernel weight tensor, are the elements of the vectors respectively; the subscript represents the rd in the output channels of the two-dimensional convolutional kernel, the subscript represents the th in the input channels of the two-dimensional convolutional kernel, and the subscript is the th in the convolutional kernel height or convolutional kernel width dimension.

[0048] When decomposing high-order tensors, the reasonable selection of the rank has always been a difficult point. Most of the existing technical methods for rank selection simply choose the decomposition rank from the perspective of the decomposition effect. Starting from the uniqueness of high-order tensor rank decomposition, the present invention constrains the rank selection within the balance of the theoretical framework and actual performance requirements. Specifically, the present invention determines the decomposition rank of the convolutional kernel weight tensor in the CP decomposition according to the Kruskal uniqueness theory and the computing power of the low-rank neural network operating device.

[0049] According to the Kruskal uniqueness theory, for a factor matrix A, the k-rank is denoted as , which is the maximum value that satisfies the condition that any column vectors of A are linearly independent. Then, for an N-order tensor, the sufficient condition for its decomposition uniqueness is: (2)(2) The convolutional kernel weight tensor is a 4th-order tensor, and its decomposition rank is denoted as . If no restriction is imposed on the rank , assuming that in the ideal case, the neural network has learned the full-rank factor matrix parameters, then the convolutional kernel decomposition rank determined according to the Kruskal uniqueness theory satisfies the following conditions: (3)(3) Considering the powerful learning ability of the neural network, the necessary condition for general decomposition uniqueness has shape restrictions for 4th-order tensors. However, in the process of neural network gradient calculation or parameter update, the requirement for consistency is relatively high, while the requirement for shape is relatively low. Secondly, there are two advantages in only considering the sufficient condition: one is to constrain the rank selection within the theoretical framework, avoiding the blindness and redundancy of rank selection; the other is to retain a certain space for rank selection, giving full play to the flexibility of the neural network's ability to learn parameters in exploring the low-rank tensor space. At the same time, considering that the selection of a high rank leads to poor compression effects, more information redundancy, and greater computational overhead, the maximum value of the rank is restricted. The maximum rank is a hyperparameter, and the maximum rank is determined according to the computing power of the low-rank neural network operating device; the lower the computing power of the low-rank neural network operating device, the higher the compression ratio required, the smaller the selection; the higher the computing power of the low-rank neural network operating device, the larger the selection; Finally, the minimum value of the decomposition rank and the maximum rank (4) (4) The forward propagation process of the low-rank convolution module is asFigure 3 As shown in the figure. For the convenience of display, assume that the input is a single 4×4 color picture, the convolution sum size is 2×2, and the output is a single channel.

[0050] At this time is 1, the default weight parameter of the first dimension is 0, and this item is omitted in the following derivation. The input of the low-rank convolution module is X, the output is L, and the forward calculation formula for each output channel is: (5) Among them, is the element in the output matrix L in the i th j row and th column of the current output channel;

[0051] Taking the gradient of the weight W, we have: (6) Assume that the loss function is E and the output feature map matrix is F. Then the gradient of the loss function with respect to W is: (7) The gradient of the loss function with respect to the elements of the factor matrix is: (8) (9) (10) (11) Assume that the learning rate is , and the gradient update formula for the factor matrix is as follows: (12) (13) (14) (15) (2) Low-rank fully connected module (LRCPFc) In classification or recognition tasks, due to the high input and output dimensions, the fully connected layer usually contains a large number of parameters, and the model calculation amount is large. The weight dimension of the fully connected layer is ( , ), which is a matrix structure. The present invention does not consider the mathematical requirements for the form of the pre-trained model parameter matrix, and directly uses the matrix rank decomposition algorithm to perform matrix decomposition on the weight matrix of the fully connected layer of the fully connected module, constructs a low-rank fully connected module, and enables the network to automatically learn low-rank information representation without relying on the weight matrix information of the pre-trained model.

[0052] The low-rank fully connected module is as follows Figure 4 shown. The fully connected weight matrix is decomposed into two factor matrices U and V, where , is the weight matrix of the fully connected layer, , , , for performing low-rank decomposition, there is: (16) where, is the element in the weight matrix of the fully connected layer , are the elements in the factor matrices respectively; the subscript represents the th in the output feature dimension of the fully connected module, and the subscript represents the th in the input feature dimension of the fully connected module; the subscript represents the th rank-1 vector in the factor matrix , is the decomposition rank of the weight matrix of the fully connected layer.

[0053] Similarly, after getting rid of the weight constraints of the pre-trained model, it is possible to no longer consider the limitation of using the SVD algorithm for non-square matrix decomposition, simplify the decomposition, and at the same time, according to the decomposition formula, there is: (17) The decomposition rank of is denoted as . If no restrictions are imposed on the rank (18) In the actual classification task, the output feature dimension , to ensure that the model still maintains good performance after decomposition and allows the network to learn more useful information, let , there is , according to formulas (17) and (18), the decomposed weight matrix satisfies . In summary, select as the matrix decomposition rank of the fully connected layer to learn and store as many useful parameters as possible while compressing a large number of parameters.

[0054] The forward process of the low-rank fully connected network is as follows Figure 5As shown. The input of the low-rank fully connected module is , and the output is ; The forward calculation formula (without considering the bias term) is: (19) Let the loss function for training the low-rank fully connected module be The gradient of the output is ; Then the gradient calculation of the weight matrix is: (20) The gradients of the factor matrices U and V are: (21) (22) Assume the learning rate is , and the gradient update formula for the factor matrices is as follows: (23) (24) where, " " represents the gradient update operation.

[0055] Adopt the low-rank convolution module and the low-rank fully connected module to replace the convolution module and the fully connected module in the convolutional neural network to obtain a low-rank neural network.

[0056] Step 2: Train the low-rank neural network: Without pre-training the network weights, use the training samples to train the low-rank neural network.

[0057] Different from most methods that require pre-training the high-rank information of the model, the present invention initializes the parameters and allows the neural network to learn the effective information of the low-rank subspace from scratch. Learning from scratch only requires initializing the network parameters of the low-rank neural network; starting from the initialized network parameters as the training starting point, train the low-rank neural network without pre-training the network weights.

[0058] In a preferred solution, the Kaiming initialization (He Kaiming initialization) is adopted to initialize the network parameters.

[0059] Step 3: Collect the target feature parameters and input them into the trained low-rank neural network, and the low-rank neural network outputs the target recognition result.

[0060] Thus, this process ends.

[0061] Embodiment 2 Based on Embodiment 1, this embodiment provides a network structure of a low-rank neural network. As Figure 6As shown, the network includes a preliminary feature extraction module, a multi-feature fusion module, and a feature re-integration and classification module.

[0062] The preliminary feature extraction module is used to perform preliminary feature extraction on the input target feature parameters, and the extracted preliminary features are input into the multi-feature fusion module.

[0063] The multi-feature fusion module is composed of N fusion extraction networks with the same structure connected in series. Each fusion extraction network includes a shallow feature extraction channel, a deep feature extraction channel, an attention feature extraction channel, and a splicing and fusion unit. Among them: The shallow feature extraction channel has feature extraction layers, and the preliminary features are output as shallow features after passing through the shallow feature extraction channel .

[0064] The deep feature extraction channel has feature extraction layers, , and the preliminary features are output as deep features after passing through the deep feature extraction channel .

[0065] The attention feature extraction channel adopts the filter balancing attention module FBAM (Filter Balancing Attention Module) designed by the present invention, which includes a first branch, a second branch, and a layer decreasing mask module D-Mask (Layer Decreasing Mask). The first branch performs convolutional feature extraction and expansion on the preliminary features to obtain the first branch features . The second branch performs convolutional feature extraction and expansion on the preliminary features , and then is processed through a Sigmoid function layer to obtain the parameter amplitude features . The layer decreasing mask module D-Mask performs a partial zeroing operation on the deep features and outputs a feature mask . As the level of the fusion extraction network gradually deepens from 1 to N , the representativeness of the features is continuously enhanced, and the zeroing ratio of the layer decreasing mask module for the deep features decays layer by layer. The first branch features , the parameter amplitude features and the feature mask are multiplied to generate attention features .

[0066] In this embodiment, the way that the layer decreasing mask module D-Mask performs a partial zeroing operation on the deep features and outputs a feature mask is as follows: For the layer fusion extraction network, calculate the zeroing ratio : (25) Wherein, and are respectively the zeroing ratios of the set first-layer fusion extraction network and the last-layer fusion extraction network, N is the total number of layers of the fusion extraction network in the multi-feature fusion module.

[0067] According to the calculated zeroing ratio , for the deep features output by the deep feature extraction channels of the layer fusion extraction network , perform partial zeroing judgment on each element: judge whether the absolute value of the element in the deep feature is greater than or equal to the absolute value threshold of the layer, and whether the generated uniform random variable is less than . If so, set the position element value in the corresponding feature mask to 0, otherwise, set it to 1; by performing partial zeroing operations on all elements of the deep feature , generate the feature mask . The D-mask formula is expressed as: For the layer parameters , the mask value at the corresponding position in the feature mask is: (26) Where: represents the element at the position in the deep feature generated by the layer fusion extraction network, is the position serial number of the element in the four-dimensional matrix; is the absolute value threshold of the layer parameters, is the uniform random variable that controls the random zeroing decision; represents the element at the position in the feature mask of the layer fusion extraction network at the position.

[0068] The layer decreasing mask module D-Mask is only used during network training. During actual recognition, the layer decreasing mask module D-Mask is removed.

[0069] In the multi-feature fusion module, obtain shallow features , deep features and attention features After that, the three features are fused, and the fused features are output , The input features are then integrated and classified by the module

[0070] The feature integration and classification module is used to further process the fused features and output the target recognition result

[0071] It can be seen from the above network structure that the design features of the neural network of the present invention are as follows First, in the multi-feature fusion module of the present invention, in addition to the deep and shallow feature extraction channels, an attention feature extraction channel is added, and the attention feature extraction channel is designed with the Filter Balanced Attention Module (FBAM). Traditional CNN has limitations in the distribution of local information and uniform weights. The attention module enhances the model's perception ability of key features through feature selection and global modeling, while ensuring high efficiency and flexibility. Different from general attention modules, the Filter Balanced Attention Module (FBAM) designed in the present invention extracts and expands features through multiple convolutional layers, and obtains the parameter amplitude features through the sigmoid function layer of the second branch. Different from the general idea of only relying on the amplitude size as the determination of parameter importance, the present invention believes that parameters with smaller amplitudes are not necessarily unimportant, and parameters with larger amplitudes do not necessarily all contain valid information. Therefore, based on this idea, a layer-by-layer decreasing mask module (D-mask) is designed. A certain proportion of the parameters with larger amplitudes in the shallow network are randomly set to zero to prevent overfitting by increasing parameter sparsity. As the network layer deepens, the feature representativeness is continuously enhanced, and the zeroing ratio gradually decreases layer by layer, retaining more high-semantic features, making the model more flexible in mining feature information and avoiding the fitting problem caused by parameter homogenization

[0072] In addition, as the number of layers of the traditional CNN model increases, feature representation deviation often occurs, and the model effect even degrades. To solve this problem, the present model proposes a method of multi-feature fusion of the model, which fuses the shallow features, deep features and attention features of the model to enhance the feature expression ability and effectively avoid the problem of biased feature extraction in a single branch

[0073] Figure 7 shows Figure 6 a specific structural example of a low-rank neural network. As shown in the figure The preliminary feature extraction module consists of a low-rank convolution module and a ReLU activation layer In the multi-feature fusion module, the shallow feature extraction channel consists of a low-rank convolution module and a batch normalization module, generating shallow features 。The deep feature extraction channel consists of a low-rank convolution module and a batch normalization module to form a set of processing units. Two sets of processing units are connected in series through a ReLU activation layer, and the batch normalization module of the second set of processing units generates deep features. 。The first branch of the attention feature extraction channel includes two low-rank convolution modules, and the second branch includes two low-rank convolution modules and a Sigmoid function layer. The outputs of the first branch, the second branch, and the D-mask module are connected to a multiplication module, and the output of the multiplication module passes through a batch normalization module to generate attention features. 。Shallow features 、Deep features And attention features Enter the splicing and fusion unit for feature fusion, and output the fused features. 。

[0074] The feature reintegration and classification module is composed of a low-rank convolution module, a batch normalization module, and a low-rank fully connected module connected in series.

[0075] Example 3 In this embodiment, based on the low-rank neural network of Example 2, a filter-balanced attention distillation network LFBADN for training the low-rank neural network is designed.

[0076] As Figure 8 shown, on the left side of the filter-balanced attention distillation network LFBADN is the low-rank neural network of Example 2, and on the right side is a homogeneous convolutional neural network in which the low-rank convolution module and the low-rank fully connected module are replaced by ordinary convolution modules and fully connected modules. The lower side of the filter-balanced attention distillation network LFBADN corresponds to the modules for soft loss calculation, hard loss calculation, and weighting operation.

[0077] Considering the low-rank neural network of Example 2, training from scratch will face the problems of uncertain training parameter directions, slow convergence, and easy convergence to local minima. Therefore, in this embodiment, to reduce the computing power pressure of model training and ensure the high quality and stability of the teacher model, the idea of offline distillation is adopted, that is, first train a homogeneous convolutional neural network as the teacher model, and then form a distillation learning network with the trained teacher model and the low-rank neural network to quickly guide the student model to learn the parameters of effective representations.

[0078] Calculate the hard loss between the output of the low-rank neural network and the true label ,Calculate the soft loss between the output of the low-rank neural network and the output of the teacher model ,The total loss of the model Is composed of the weighted sum of the two, that is: (27) Among them, Is the weighting coefficient for weighted calculation of the hard loss and the soft loss.

[0079] Parameter optimization of a low-rank neural network based on the total loss.

[0080] The low-rank filter balanced attention distillation network LFBADN of this embodiment enhances the perception ability of key features through the filter balanced attention module FBAD, fuses multi-features to enhance the feature expression ability, and guides the low-rank neural network to quickly learn effective parameters through the isomorphic teacher model with a large number of parameters, realizing accurate target recognition. The main advantages of this model are as follows: ① Anti-training degradation ability: Multi-feature fusion can effectively avoid the problems of gradient vanishing and gradient explosion, thereby improving the training stability and robustness of the network.

[0081] ② Simple training: Different from general low-rank decomposition that requires a large number of iterative fine-tuning, this model designs a low-rank module and trains from scratch, which is simple and efficient.

[0082] ③ Efficient feature extraction: Design a filter balanced attention module, synthesize the overall features of the filter, focus on key features, and at the same time use hierarchical masking to improve the anti-interference ability of the model and prevent the model from overfitting.

[0083] ④ Distillation design: By designing the distillation structure of the low-rank network and the isomorphic non-low-rank network, use the teacher model with more parameters to guide the low-rank network to learn effective parameters from scratch, and the training is more likely to converge and the accuracy is higher.

[0084] ⑤ Scalability: The low-rank module is not only applicable to this model, but can be extended to other deep learning tasks with similar model structures.

[0085] Next, actual data is used to train the model of this embodiment, and the training process and results are as follows: (1) Dataset construction The SAR image dataset used in this invention contains four categories: Cargo, Tanker, Other Type, and Tug. These four target categories are divided into training data and test data, and the division ratio is 7:3. Each category in the training set contains 1610 SAR images for the model to learn and train; each category in the test set contains 690 SAR images for evaluating the performance of the model on unknown data. Table 1 shows the classification of the dataset.

[0086] Table 1 Dataset classification

[0087] (2) Training and validation of the model Next, the model and training parameter settings will be described.

[0088] ① The idea of the lightweight network designed in the present invention is to train a low-rank network from scratch, rather than decomposing the weight parameters of a pre-trained model. The low-rank decomposition algorithms based on pre-trained models often require iterative fine-tuning, consuming a large amount of time and computing power, and the accuracy drops severely when the compression ratio is high. For the decomposition method proposed in the present invention, for CP decomposition, due to the problem of choosing the CP rank, if the same decomposition is performed on the pre-trained model, it will lead to insufficient representation of parameter information under the same decomposition effect, and the performance drops more severely; for matrix decomposition, the decomposition method proposed in the present invention does not satisfy the mathematical derivation of general matrix decomposition (such as SVD decomposition), and using a pre-trained model for such matrix decomposition cannot effectively obtain the high-precision relevant information in the pre-trained parameters, and the decomposition effect is poor. In summary, the present invention adopts the idea of training from scratch, enabling the network to obtain effective parameters under a given rank during learning. Therefore, a parameter initialization strategy needs to be adopted. The network training of the present invention uses Kaiming initialization to better adapt to the gradient problem caused by parameter decomposition.

[0089] ② The batch_size is set to 16, the number of training epochs is set to 50, and single-GPU training and inference are performed.

[0090] ③ The operating device parameters are as follows: CPU: AMD Ryzen 7 4800H with Radeon Graphics 2.90 GHz; GPU: GeForce RTX 2060 6G, and the PyTorch deep learning framework is used.

[0091] In this SAR target recognition model, the number of target categories is 4. According to the principle derivation, the decomposition rank of the low-rank fully connected layer is set to 4, and different maximum decomposition ranks are set for the low-rank convolution module. The number of model parameters and recognition rate under different compression conditions are shown in Table 2.

[0092] Table 2 Verification Results

[0093] Example 4 This embodiment provides an object recognition device based on a low-rank neural network. As Figure 9 shown, it includes a low-rank neural network and a training module.

[0094] Among them, the low-rank neural network is obtained by replacing the convolution module in the convolutional neural network with a low-rank convolution module and the fully connected module with a low-rank fully connected module; the collected target feature parameters are input into the low-rank neural network, and the low-rank neural network outputs the target recognition result.

[0095] The determination method of the low-rank convolution module is as follows: perform CP decomposition on the convolution kernel weight tensor of the two-dimensional convolution kernel in the convolution module; determine the decomposition rank of the convolution kernel weight tensor in the CP decomposition according to the Kruskal uniqueness theory and the computing power of the low-rank neural network operating device; obtain the low-rank convolution module through CP decomposition.

[0096] The determination method of the low-rank fully connected module is as follows: perform matrix decomposition on the weight matrix of the fully connected layer of the fully connected module; determine the decomposition rank of the weight matrix of the fully connected layer according to the input feature dimension and output feature dimension of the fully connected module; obtain the low-rank fully connected module through matrix decomposition.

[0097] The training module is used to train the low-rank neural network with training samples without pre-training the network weights.

[0098] The low-rank neural network in this embodiment can adopt the specific network structure of Embodiment 2. The training method can refer to the distillation training scheme of Embodiment 3.

[0099] Compared with the prior art, the solution of the present invention has the following advantages: 1. Analyzed from the compression effect and the performance after model compression. For the traditional low-rank decomposition of the high-dimensional weight tensor of the pre-trained model, there are convergence problems, the training cost is relatively high, and the performance drops severely when the compression ratio is relatively high. For the traditional low-rank decomposition of the fully connected layer, the singular value decomposition (SVD) algorithm is mostly used, which brings greater computational complexity. While achieving a good effect in model compression, it often brings a large performance loss. The present invention has a very small loss of model performance while maintaining a great compression effect, and the training is simple.

[0100] 2. Analyzed from the selection of the decomposition rank. For the selection of the decomposition rank by traditional other algorithms, especially for the low-rank decomposition of high-order tensors, it mostly starts from artificial experience. The present invention starts from the uniqueness theory and the actual compression requirements in the rank decomposition of high-order tensors, rationalizes the selection of the rank, and also derives the selection of the matrix decomposition rank mathematically, combined with the characteristics of the actual data, so as to select a reasonable rank.

[0101] 3. Analyzed from the computational overhead. In the traditional low-rank decomposition of the pre-trained model, multiple iterations of fine-tuning are required, the computational overhead is relatively large, and it is easy to converge to the local minimum point. The present invention trains the low-rank module from scratch, converges quickly, and has a relatively small computational overhead.

[0102] 4. Analyzed from the decomposition integrity. Most traditional algorithms only decompose a certain module of the neural network. The present invention designs decomposition algorithms respectively according to the characteristics of different modules, constructs different low-rank modules, and has strong decomposition integrity.

[0103] 5. Analysis from the aspect of scalability. Most of the traditional other low-rank decomposition algorithms are designed for a specific neural network, while the present invention designs a basic network module, which has good scalability.

[0104] The above specific embodiments only describe the design principle of the present invention. The shapes and names of the components in this description can be different and are not limited. Therefore, those skilled in the art of the present invention can modify or equivalently replace the technical solutions recorded in the foregoing embodiments; and these modifications and replacements do not depart from the purpose and technical solutions of the present invention, and shall fall within the protection scope of the present invention.

Claims

1. A target recognition method based on a low-rank neural network, characterized in that, Including: Step 1: Design a low-rank neural network: Design a low-rank convolutional module: Perform CP decomposition on the convolutional kernel weight tensor of the two-dimensional convolutional kernel in the convolutional module; Determine the decomposition rank of the convolutional kernel weight tensor in the CP decomposition according to the Kruskal uniqueness theory and the computing power of the device on which the low-rank neural network runs; Obtain the low-rank convolutional module through CP decomposition; Design a low-rank fully connected module: Perform matrix decomposition on the weight matrix of the fully connected layer of the fully connected module; Determine the decomposition rank of the weight matrix of the fully connected layer according to the input feature dimension and output feature dimension of the fully connected module; Obtain the low-rank fully connected module through matrix decomposition; Use the low-rank convolutional module and the low-rank fully connected module to replace the convolutional module and the fully connected module in the convolutional neural network to obtain a low-rank neural network; Step 2: Train the low-rank neural network: Without pre-training the network weights, use training samples to train the low-rank neural network; Step 3: Collect target feature parameters and input them into the trained low-rank neural network, and the low-rank neural network outputs the target recognition result.

2. The object recognition method based on a low-rank neural network according to claim 1, characterized in that The CP decomposition of the convolutional kernel weight tensor of the two-dimensional convolutional kernel in the convolutional module is: Decompose the convolutional kernel weight tensor CP into factor matrices , , , , where , , , are the rank-1 vectors that make up the factor matrices , , , , respectively; , is the decomposition rank of the convolutional kernel weight tensor; The four dimensions of the two-dimensional convolutional kernel are the output channel dimension, the input channel dimension, the convolutional kernel height, and the convolutional kernel width, and the convolutional kernel height and the convolutional kernel width are the same; The convolutional kernel weight tensor is denoted as , and for performing CP decomposition, we have: Among them, " " represents the vector cross product operation; is an element in the convolutional kernel weight tensor , are respectively elements in the vector ; the subscript represents the -th in the number of output channels of the two-dimensional convolutional kernel, and the subscript represents the -th in the number of input channels of the two-dimensional convolutional kernel, and the subscript is the -th in the convolutional kernel height or convolutional kernel width dimension; The matrix decomposition of the weight matrix of the fully connected layer of the fully connected module is: The weight matrix of the fully connected layer is respectively the factor matrix U and the factor matrix V; Let be the weight matrix of the fully connected layer. For matrix factorization, we have: Among them, is an element of the fully connected layer weight matrix ; are elements of the factor matrix respectively; The subscript represents the -th in the output feature dimension of the fully connected module, and the subscript represents the -th in the input feature dimension of the fully connected module; The subscript r represents the -th rank-1 vector in the factor matrix , is the decomposition rank of the fully connected layer weight matrix.

3. The object recognition method based on a low-rank neural network according to claim 1, characterized in that The determination of the decomposition rank of the convolutional kernel weight tensor in the CP decomposition according to the Kruskal uniqueness theory and the computing power of the device on which the low-rank neural network runs is: Determine the first decomposition rank according to the Kruskal uniqueness theory ; Determine the maximum rank according to the computing power of the low-rank neural network running device ; The lower the computing power of the low-rank neural network running device, the higher the compression ratio required, the smaller the selection; the higher the computing power of the low-rank neural network running device, then the larger the selection; With the first decomposition rank and the maximum rank The minimum value is used as the decomposition rank of the convolutional kernel weight tensor.

4. The object recognition method based on a low-rank neural network according to claim 3, wherein The first decomposition rank determined according to the Kruskal uniqueness theory is as follows: The convolution kernel weight tensor is a fourth-order tensor, and the first decomposition rank is deduced according to the Kruskal uniqueness theory Satisfy the following conditions: Among them, is the number of output channels of the two-dimensional convolution kernel; is the number of input channels of the two-dimensional convolution kernel; is the height and width of the convolution kernel in the two-dimensional convolution kernel, and the height is the same as the width.

5. The object recognition method based on a low-rank neural network according to claim 1, characterized in that The determination of the decomposition rank of the weight matrix of the fully connected layer according to the input feature dimension and output feature dimension of the fully connected module is: The decomposition rank of the weight matrix of the fully connected layer takes the output feature dimension of the fully connected module.

6. The object recognition method based on a low-rank neural network according to claim 1, characterized in that, In Step 2, during the training process, use Kaiming to initialize the network parameters of the low-rank neural network; Starting from the initialized network parameters, without pre-training the network weights, train the low-rank neural network.

7. The object recognition method based on a low-rank neural network according to claim 1, characterized in that The low-rank neural network includes a preliminary feature extraction module, a multi-feature fusion module, and a feature re-integration and classification module; The preliminary feature extraction module is used to perform preliminary feature extraction on the input target feature parameters, and the extracted preliminary features are input into the multi-feature fusion module; The multi-feature fusion module is composed of N fusion extraction networks with the same structure connected in series; each fusion extraction network includes a shallow feature extraction channel, a deep feature extraction channel, an attention feature extraction channel, and a splicing and fusion unit; the shallow feature extraction channel has feature extraction layers, and the preliminary features pass through the shallow feature extraction channel to output shallow features ; The deep feature extraction channel has feature extraction layers, , and the preliminary features are output as deep features through the deep feature extraction channel ; In the attention feature extraction channel: The first branch performs convolutional feature extraction and expansion on the preliminary features to obtain the first branch features ; The second branch performs convolutional feature extraction and expansion on the preliminary features and then processes them through a Sigmoid function layer to obtain the parameter amplitude features ; The layer-decreasing masking module D-Mask performs a partial zeroing operation on the deep features and outputs a feature mask ; As the level of the fusion extraction network gradually deepens from 1 to N , the representativeness of the features continuously strengthens, and the zeroing ratio of the layer-decreasing masking module for the deep features decays layer by layer; The first branch features , the parameter amplitude features and the feature mask are multiplied to generate the attention features ; The layer-decreasing masking module D-Mask is only used during network training; The shallow features , deep features and attention features are fused to output the fused features , which are input into the feature re-integration and classification module; The feature re-integration and classification module processes the fused features and outputs the target recognition result.

8. The object recognition method based on a low-rank neural network according to claim 7, wherein The layer decreasing mask module D-Mask performs partial zeroing on the deep features and outputs a feature mask in the following manner: For N the layer in the layer fusion extraction network, calculate the zeroing ratio : Among them, and are the zeroing ratios of the set first-layer fusion extraction network and the last-layer fusion extraction network respectively; According to the reset ratio , for the deep feature extraction channels of the deep layer fusion extraction network output deep features perform partial zeroing judgment on each element in: judge whether the absolute value of the element in the deep feature is greater than or equal to the absolute value threshold set for the layer; and generate a uniform random variable, judge whether the uniform random variable is less than ; if both the judgment of the absolute value of the element and the uniform random variable are yes, set the position of the current element in the feature mask to 0, otherwise, set it to 1; by performing partial zeroing judgment operations on all elements of the deep feature , generate a feature mask .

9. The object recognition method based on a low-rank neural network according to claim 7, characterized in that When training the low-rank neural network, construct an isomorphic convolutional neural network that does not replace the low-rank convolutional module and the low-rank fully connected module, first train the isomorphic convolutional neural network as a teacher model, and then form a distillation learning network with the trained teacher model and the low-rank neural network to perform distillation learning on the low-rank neural network.

10. An object recognition device based on a low-rank neural network, characterized in that, The device includes a low-rank neural network and a training module; The low-rank neural network is obtained by replacing the convolutional module in the convolutional neural network with a low-rank convolutional module and the fully connected module with a low-rank fully connected module; The collected target feature parameters are input into the low-rank neural network, and the low-rank neural network outputs the target recognition result; The determination method of the low-rank convolution module is as follows: perform CP decomposition on the convolution kernel weight tensor of the two-dimensional convolution kernel in the convolution module; determine the decomposition rank of the convolution kernel weight tensor in the CP decomposition according to the Kruskal uniqueness theory and the computing power of the low-rank neural network operating device; obtain the low-rank convolution module through CP decomposition; The determination method of the low-rank fully connected module is as follows: perform matrix decomposition on the weight matrix of the fully connected layer of the fully connected module; determine the decomposition rank of the weight matrix of the fully connected layer according to the input feature dimension and output feature dimension of the fully connected module; obtain the low-rank fully connected module through matrix decomposition; The training module is used to train the low-rank neural network with training samples without pre-training the network weights.

Images