FPGA-based deep feature fusion network narrowband radar target recognition system

By deploying an accelerator IP core of a deep feature fusion network on an FPGA platform and utilizing the programmable logic of the FPGA to implement weight caching and ping-pong buffering, the problems of high power consumption and long latency in narrowband radar target recognition systems are solved, achieving miniaturized, low-cost, and efficient recognition.

CN117828526BActive Publication Date: 2026-11-10XIDIAN UNIV
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202311817265.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-12-27
Publication Date
2026-11-10
Estimated Expiration
2043-12-27

AI Technical Summary

Technical Problem

Existing deep learning-based narrowband radar target recognition methods suffer from high power consumption and time delays on hardware platforms, which are also large in size and expensive.

Method used

An FPGA hardware platform is used to deploy an accelerator IP core for a deep feature fusion network. The programmable logic (PL) terminal of the FPGA is used to implement weight caching and ping-pong buffering to reduce the frequency of memory access. The computational efficiency is improved by one-dimensional convolution and fully connected acceleration modules.

Benefits of technology

It effectively reduces system power consumption and latency while ensuring the accuracy and reliability of identification. The hardware platform is small in size and low in cost.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117828526B_ABST
    Figure CN117828526B_ABST
Patent Text Reader

Abstract

This invention proposes a narrowband radar target recognition system based on an FPGA using a deep feature fusion network. The system comprises a deep feature fusion network deployed on an FPGA, including an encoding module and cascaded, parallel-arranged decoding and fully connected modules. The invention utilizes a weight caching module and a ping-pong buffer module to cache network weight data and intermediate results generated by network computation, respectively, eliminating the need for frequent access to the PS-side memory and effectively reducing system power consumption. The one-dimensional convolution acceleration module and the fully connected acceleration module enable rapid computation, and the weight caching module and the ping-pong buffer module can create different cache spaces, further reducing network latency.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of narrowband radar target recognition, and relates to a narrowband radar target recognition system, specifically a narrowband radar target recognition system based on FPGA deep feature fusion network. Background Technology

[0002] Radar target identification requires extracting relevant information and stable features from the target's radar echo to determine its attributes. Narrowband low-resolution radar has a resolution capability in both the lateral and longitudinal dimensions that is much larger than the target size. The target is approximated as a "point" target, making it impossible to extract the fine structure of the target. However, it can extract many features at the level of unique attributes. These features are an important support for narrowband radar target identification.

[0003] Narrowband radar target recognition is widely used in the rapid identification of aerial maneuvering targets. Due to the high speed and maneuverability of aerial maneuvering targets, as well as the limited storage space of the hardware platform in this scenario and the requirement for small size and low power consumption, the narrowband radar target recognition system is required to be miniaturized, have low latency, and low power consumption, in addition to ensuring the accuracy and reliability of the system.

[0004] Traditional narrowband radar target recognition methods often employ statistical analysis to construct probabilistic models, then directly fuse shallow features such as time-domain, frequency-domain, and Doppler-domain features based on these models, finally classifying the target based on the fused features. However, these methods typically rely on simple statistical analysis to directly combine shallow features, failing to extract more representative deep features. Furthermore, the separability of shallow features is limited, making it difficult to comprehensively reflect the target's intrinsic information, thus limiting recognition accuracy. Deep learning-based narrowband radar target recognition methods can handle more complex situations with less overhead, possessing powerful deep feature representation capabilities that can significantly improve recognition accuracy. However, compared to traditional methods, deep learning-based methods require substantial data storage and computational overhead, thus placing higher demands on hardware platform performance during practical deployment.

[0005] For example, application CN117079034A, entitled "A Classification Method for Narrowband Radar Aerial Targets Based on Deep Feature Fusion Network," discloses a narrowband radar target recognition system. This system includes a deep feature fusion network and a hardware platform deploying the network using a CPU + GPU architecture. The deep feature fusion network includes an encoding module, a cascaded and parallel-arranged decoding module, a fully connected module, and a Softmax function layer cascaded at the output of the fully connected module. This layer is used to identify aerial targets, specifically including the identification of targets outside the target pool and the classification of targets within the target pool. During the training of the network and the acquisition of classification results, the encoding module is used to obtain deep, highly separable features, effectively improving the classification accuracy. The decoding module is used to reconstruct the input features and obtain the target reconstruction error, enabling the network to identify targets outside the target pool and improving the reliability of classification. The CPU is used for system control, and the GPU is used to accelerate network inference and reduce system latency. However, its drawbacks are that during the operation of the deep feature fusion network, the GPU needs to frequently read and write to the external memory of the CPU, which consumes a lot of energy and results in high system power consumption. In addition, the GPU needs to wait for batch alignment before it can perform inference on the network, which results in high system latency. At the same time, the hardware platform of this invention is large in size and expensive. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the existing technology and propose a narrowband radar target recognition system based on FPGA deep feature fusion network. This system aims to ensure classification accuracy and reliability while effectively reducing system power consumption and latency.

[0007] To achieve the above objectives, the technical solution adopted by the present invention includes a deep feature fusion network deployed on a hardware platform, wherein the hardware platform adopts an FPGA including a processor system (PS) and a programmable logic unit (PL); the deep feature fusion network is an accelerator IP core located at the PL, including an encoding module unit and a decoding module unit and a fully connected module unit cascaded and arranged in parallel therewith, as well as an AXI interface module located between the encoding module unit and the decoding module unit and the fully connected module unit; wherein:

[0008] The encoding module unit includes a cascaded input feature buffer module and multiple encoding modules. Each encoding module includes a cascaded one-dimensional convolution acceleration module, a pooling module, and an attention mechanism module, each with its output end connected to a ping-pong buffer module. The input ends of the one-dimensional convolution acceleration module and the attention mechanism module are also each connected to a weight buffer module.

[0009] The decoding module unit includes multiple cascaded decoding modules and a reconstruction error module. Each decoding module includes a cascaded upsampling module and a one-dimensional convolution acceleration module, each with its output connected to a ping-pong buffer module. The input of the one-dimensional convolution acceleration module is also connected to a weight cache module.

[0010] The fully connected module unit includes multiple cascaded fully connected modules. Each fully connected module includes a cascaded ping-pong buffer module and a fully connected acceleration module. The input end of the fully connected acceleration module is also connected to a weight cache module.

[0011] As an optimization, the processor system-side PS includes a processor and a memory that interacts with it; the processor is used to preprocess narrowband radar feature data received through the UART serial port, control the state of the accelerator IP core, postprocess the output results of the accelerator IP core, and output the processing results through the UART serial port; the memory is used to cache the weight data of the deep feature fusion network and the processing data of the processor.

[0012] As an optimization, the encoding module unit, wherein:

[0013] The input feature caching module is used to cache the narrowband radar feature data input from the AXI interface module;

[0014] The one-dimensional convolution acceleration module is used to perform fast one-dimensional convolution and activation operations on the features input by the encoding module. It includes a cascaded padding module, a DSP array module, and an activation calculation module, as well as weight and bias units connected to the DSP array module. The padding module is used to zero-padded the edges of the input features. The DSP array module includes M×N DSP units, used to perform parallel multiplication and addition operations on the features output by the padding module and the values ​​in the weight and bias units. The activation calculation module is used to perform ReLU or PReLU activation calculations on the output of the DSP array module. M and N represent the parallelism in the output feature size dimension and the parallelism in the output feature channel dimension, respectively.

[0015] The pooling module is used to perform max pooling calculation on the features output by the ping-pong buffer module connected to its input terminal.

[0016] The attention mechanism module is used to extract multi-dimensional features from the features output by the ping-pong buffer module cascaded at the output of the pooling module. It includes a cascaded spatial attention mechanism module and a channel attention mechanism module. The spatial attention mechanism includes two pooling modules arranged in parallel and a fully connected acceleration module cascaded at their outputs. The two pooling modules are used for global average pooling and global max pooling calculations, respectively. The channel attention mechanism module includes two pooling modules arranged in parallel and a one-dimensional convolution acceleration module cascaded at their outputs. The two pooling modules are used for global average pooling and global max pooling calculations, respectively.

[0017] The ping-pong buffer module is used to buffer the data output by the one-dimensional convolution acceleration module, pooling module, or attention mechanism module connected to its input end.

[0018] The weight cache module is created using BRAM resources within the programmable logic terminal and is used to cache the weight parameters in the accelerator IP core.

[0019] As an optimization, the size of the DSP array module is determined by the parameters of the port of the one-dimensional convolution acceleration module.

[0020] As an optimization, the decoding module unit, wherein:

[0021] The reconstruction error module is used to perform mean square error statistics on the reconstructed feature data output by the decoding module connected to its input terminal and the original features input by the AXI interface module in order to obtain the reconstruction error.

[0022] The upsampling module is used to upsample the input features of the decoding module.

[0023] As an optimization, the fully connected module unit, wherein the fully connected acceleration module is used to perform fast fully connected and activation operations on the features output by the ping-pong buffer module, the fully connected acceleration module includes cascaded DSP computing units and activation computing modules, as well as weight units and bias units connected to the DSP computing units; the DSP computing unit includes N DSP units, which are used to perform parallel multiply-add operations on the features input to the fully connected acceleration module and the values ​​in the weight units and bias units, and the activation computing module is used to perform ReLU activation calculations on the output of the DSP computing units; where N represents the parallelism of the output feature size dimension.

[0024] As an optimization, the size of the DSP computing unit is determined by the parameters of the fully connected acceleration module port.

[0025] As an optimization, the AXI interface module is used to transmit processing information of the encoding module unit, decoding module unit, and fully connected module unit in the processor and accelerator IP core, and to store the status and function information of the encoding module unit, decoding module unit, and fully connected module unit.

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

[0027] 1. The deep feature fusion network in this invention is deployed at the programmable logic (PL) end of the FPGA. During the target recognition process, the weight cache module and ping-pong buffer module created by the BRAM resources of the PL respectively realize the caching of network weight data and intermediate results generated by network operation. There is no need to frequently access the memory of the PS end, which effectively reduces the system power consumption compared with the prior art.

[0028] 2. This invention uses a one-dimensional convolution acceleration module to perform fast one-dimensional convolution and activation operations on the features input to the encoding module, and a fully connected acceleration module to perform fast fully connected operations and activation operations on the features output by the ping-pong buffer module. A weight caching module creates a weight parameter cache space around the functional modules, reducing the time spent reading and writing weight parameter data. The ping-pong buffer module creates independent read / write buffers between different functional modules, enabling the accelerator IP core to have multi-stage pipeline capabilities, avoiding the high latency issue in existing technologies caused by the GPU needing to wait for batch alignment before performing inference on the network.

[0029] 3. The FPGA hardware platform used in this invention has the advantages of small size and low cost compared with the CPU+GPU architecture hardware platform used in the prior art.

[0030] 4. In the accelerator IP core of the present invention, the encoding module unit extracts features from the input step by step through multiple encoding modules to obtain highly separable target feature information, thereby ensuring the classification accuracy of the system; the decoding module unit decodes the features output by the encoding module unit through multiple decoding modules, and obtains the reconstruction error through the reconstruction error module, enabling the system to have the ability to identify targets outside the database, thereby ensuring the classification reliability of the system. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the overall structure of the present invention;

[0032] Figure 2 This is a schematic diagram of the structure of the encoding module unit of the present invention;

[0033] Figure 3 This is a schematic diagram of the structure of the encoding module of the present invention;

[0034] Figure 4This is a schematic diagram of the structure of the decoding module unit of the present invention;

[0035] Figure 5 This is a schematic diagram of the decoding module of the present invention;

[0036] Figure 6 This is a schematic diagram of the structure of the fully connected module of the present invention. Detailed Implementation

[0037] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0038] Reference Figure 1 This invention includes a deep feature fusion network deployed on a hardware platform, wherein the hardware platform employs an FPGA comprising a processor system (PS) and a programmable logic unit (PL); the deep feature fusion network is an accelerator IP core located at the PL, comprising an encoding module unit and cascaded and paralleled decoding module units and fully connected module units, and an AXI interface module disposed between the encoding module unit and the decoding module unit and the fully connected module units; wherein:

[0039] The processor system terminal (PS) includes a processor and a memory that interacts with it; the processor is used to preprocess narrowband radar feature data received via UART serial port, control the state of accelerator IP core, postprocess the output results of accelerator IP core, and output the processing results via UART serial port; the memory is used to cache the weight data of the deep feature fusion network and the processing data of the processor.

[0040] In this embodiment, the processor uses an ARM Cortex-A9, the memory uses DDR3, and the narrowband radar feature data comes from aerial targets, covering target scattering characteristics, motion characteristics, and aircraft characteristics. It consists of time series data from six channels with different features, including radial range, pitch angle, heading angle, radial velocity, active radar RCS, and semi-active radar RCS. The processor's preprocessing operation on the narrowband radar feature data specifically involves performing Z-score normalization on the feature sequences of the four channels: radial velocity, radial range, heading angle, and pitch angle. The principle of Z-score normalization is as follows: Where σ is the standard deviation, X std To normalize the radial velocity, radial range, heading angle, and pitch angle feature sequences after standardization, L2 norm normalization is performed on the feature sequences of the active radar RCS and semi-active radar RCS channels. The principle of L2 norm normalization is as follows: Where ||X|| is the L2 norm of sequence X, X norm The normalized active RCS sequence and semi-active RCS sequence are given; the accelerator IP core output includes a reconstruction error L.r And the classification probability P, the post-processing of the accelerator IP core output is as follows: First, determine L r Does it satisfy L with the preset threshold τ? r If the value is greater than or equal to τ, it is determined to be an external target and is removed from the database. Otherwise, the target is determined to be an internal target and is then classified. The category corresponding to the maximum value in P is selected as the classification result of the target. In this embodiment, τ = 10.

[0041] Reference Figure 2 and 3 The encoding module unit includes a cascaded input feature buffer module and first, second, and third encoding modules E1, E2, and E3. Each encoding module includes a cascaded one-dimensional convolution acceleration module, a pooling module, and an attention mechanism module, each with its output connected to a ping-pong buffer module. The inputs of the one-dimensional convolution acceleration module and the attention mechanism module are also connected to a weight buffer module. The encoding module unit extracts features from the input step by step through the three encoding modules to obtain highly separable target feature information, thereby ensuring the classification accuracy of the system.

[0042] The encoding module unit, wherein:

[0043] The input feature caching module is used to cache the narrowband radar feature data input from the AXI interface module;

[0044] The one-dimensional convolution acceleration module is used to perform fast one-dimensional convolution and activation operations on the features input to the encoding module. It includes a cascaded padding module, a DSP array module, and an activation calculation module, as well as weight and bias units connected to the DSP array module. The padding module performs zero-padding on the edges of the input features to ensure that the input and output feature sizes are consistent. The DSP array module includes M×N DSP units, used to perform parallel multiplication and addition operations on the features output from the padding module and the values ​​in the weight and bias units. The values ​​in the weight and bias units are derived from the weight cache module, and M and N represent the parallelism in the output feature size dimension and the parallelism in the output feature channel dimension, respectively. The activation calculation module performs ReLU or PReLU activation calculations on the output of the DSP array module.

[0045] Since deep feature fusion networks involve a large number of one-dimensional convolution operations, this invention utilizes a DSP array module within the one-dimensional convolution acceleration module to simultaneously perform M×N multiplication-accumulation operations, thereby improving system computational efficiency and effectively reducing system latency. Furthermore, the size of the DSP array module is determined by the parameters of the one-dimensional convolution acceleration module's ports. By setting these parameters, the DSP array size can be adjusted to avoid resource waste caused by mismatch between the DSP array size and the one-dimensional convolution operation scale, thus saving system power consumption. Compared to the existing technology's separate structure of convolution and activation modules, this invention integrates an activation calculation module into the one-dimensional convolution acceleration module. This structure reduces redundant data accesses between modules, further reducing system latency and power consumption. In this embodiment, considering the one-dimensional convolution operation scale in the three encoding modules, DSP arrays of 5×4, 5×4, and 4×4 scales are used sequentially for the three one-dimensional convolution acceleration modules in E1, E2, and E3, where the activation calculation modules are all used for ReLU activation calculations.

[0046] The pooling module is used to perform max pooling calculation on the features output by the ping-pong buffer module connected to its input terminal. In this embodiment, the pooling modules in E1, E2 and E3 are all used to implement max pooling calculation with a pooling window size of 1×2 and a step size of 2, so as to achieve the effect of feature dimensionality reduction.

[0047] The attention mechanism module is used to extract multi-dimensional features from the features output by the ping-pong buffer module cascaded at the output of the pooling module. It includes a cascaded spatial attention mechanism module and a channel attention mechanism module. The spatial attention mechanism includes two parallel pooling modules and a fully connected acceleration module cascaded with their outputs. The two pooling modules are used for global average pooling and global max pooling calculations, respectively. The principle is to first perform global average pooling and global max pooling calculations on each channel of the input features, then extract features through the fully connected acceleration module to obtain two different sub-features, and finally fuse the two sub-features element-wise. The channel attention mechanism module includes two parallel pooling modules and a one-dimensional convolution acceleration module cascaded with their outputs. The two pooling modules are used for global average pooling and global max pooling calculations, respectively. The principle is to first perform global average pooling and global max pooling calculations on the input features, then concatenate the two obtained sub-features, and finally fuse the key features of the channel dimension through the one-dimensional convolution acceleration module.

[0048] In this embodiment, the fully connected acceleration module DSP computing units of the first, second, and third channel attention mechanism modules C1, C2, and C3 in E1, E2, and E3 are 2, 2, and 4 respectively. Since the number of convolution parameters in the first, second, and third spatial attention mechanism modules S1, S2, and S3 in E1, E2, and E3 is only 2×7, their computational latency has a small impact on the overall system latency performance, and the acceleration of this convolution operation is not significant. Therefore, the DSP array size of the one-dimensional convolution acceleration module in S1, S2, and S3 is 1×1 to reduce system resource consumption and power consumption.

[0049] The ping-pong buffer module is used to buffer the data output by the one-dimensional convolution acceleration module, pooling module, or attention mechanism module connected to its input. It contains two buffer units, one for writing and one for reading. The size of each buffer unit is equal to the memory size occupied by the output result of the previous functional module. This module enables the accelerator IP core to have multi-stage pipeline capability. The principle is that while the subsequent functional module is reading data from one of the buffer units, the previous functional module can write its subsequent calculation results into the other buffer unit. Through this alternating storage method, seamless real-time data transmission is achieved, so that the accelerator IP core can start the next forward inference calculation before the current forward inference calculation is completed, thereby reducing the total latency of the system performing multiple recognition tasks.

[0050] In this embodiment, E1, E2, and E3 each contain three ping-pong buffer modules. The first ping-pong buffer module in E1, E2, and E3 is used to buffer the data output by the one-dimensional convolution acceleration module in E1, E2, and E3, respectively. The second ping-pong buffer module in E1, E2, and E3 is used to buffer the data output by the pooling module in E1, E2, and E3, respectively. The third ping-pong buffer module in E1, E2, and E3 is used to buffer the data output by the attention mechanism module in E1, E2, and E3, respectively.

[0051] The weight caching module, created using BRAM resources within the programmable logic, is used to cache weight parameters in the accelerator IP core. It consumes a certain amount of BRAM resources. Where N is the storage space occupied by the weight parameters, and M is a BRAM storage space in the programmable logic terminal PL, in bytes (B). This represents rounding up; after the IP core starts, it reads the weight parameters of the deep feature fusion network in advance through the AXI interface module and writes them into each weight cache module. Since the system's access latency and energy consumption to the BRAM memory inside the PL are much lower than the system's access to the PS memory, and the writing of the weight cache module can be performed synchronously with the calculation of different functional modules, the number of times the AXI interface module accesses the PS can be reduced, the system's memory access efficiency can be improved, and the system's power consumption and latency can be reduced.

[0052] In this embodiment, E1, E2 and E3 each contain two weight buffer modules. The first weight buffer module in E1, E2 and E3 is used to cache the weight parameters of the one-dimensional convolution acceleration module in E1, E2 and E3, respectively. The second weight buffer module in E1, E2 and E3 is used to cache the weight parameters of the attention mechanism module in E1, E2 and E3, respectively.

[0053] Reference Figure 4 and 5 The decoding module unit includes cascaded first, second, and third decoding modules D1, D2, and D3, as well as a reconstruction error module. Each decoding module includes a cascaded upsampling module and a one-dimensional convolution acceleration module, each with its output connected to a ping-pong buffer module. The input of the one-dimensional convolution acceleration module is also connected to a weight buffer module. The decoding module unit first decodes the features output by the encoding module unit through the three decoding modules to reconstruct the input features. Finally, it obtains the reconstruction error through the reconstruction error module, enabling the system to distinguish targets outside the database and ensuring the classification reliability of the system.

[0054] The decoding module unit, wherein:

[0055] The reconstruction error module is used to perform mean square error statistics on the reconstructed feature data obtained from D3 and the original feature data input from the AXI interface module to obtain the reconstruction error.

[0056] The upsampling module is used to upsample the input features of D1, D2 and D3. Its principle is to traverse each channel of the feature and interpolate the newly added positions according to the nearest neighbor principle, thereby increasing the size of the output feature. In this embodiment, the dimensionality increase scale of the upsampling modules in D1, D2 and D3 is 2, that is, the feature size output by each upsampling module is twice the size of the input feature.

[0057] The ping-pong buffer module in the decoding module unit is used to buffer the data output by the upsampling module and the one-dimensional convolution acceleration module. In this embodiment, there are two ping-pong buffer modules in D1, D2 and D3 respectively. The first ping-pong buffer module in D1, D2 and D3 is used to buffer the data output by the upsampling module in D1, D2 and D3 respectively. The second ping-pong buffer module in D1, D2 and D3 is used to buffer the data output by the one-dimensional convolution calculation module in D1, D2 and D3 respectively.

[0058] The one-dimensional convolution acceleration module is used in the decoding module unit to perform fast one-dimensional convolution and activation operations on the output of the upsampling module. In this embodiment, the three one-dimensional convolution acceleration modules in D1, D2 and D3 are DSP arrays of scale 5×4, 5×4 and 4×4 respectively. The activation calculation module is used for PRelu activation calculation.

[0059] In this embodiment, each of D1, D2, and D3 contains a weight buffer module, which is used to cache the weight parameters of the one-dimensional convolution acceleration module in D1, D2, and D3, respectively.

[0060] Reference Figure 6 The fully connected module unit includes cascaded first, second, and third fully connected modules L1, L2, and L3. Each fully connected module includes cascaded ping-pong buffer modules and fully connected acceleration modules. The input of the fully connected acceleration module is also connected to a weight cache module. The fully connected module unit is used to calculate the classification result based on the features extracted by the encoding module, and its output is the classification probability P, representing the probability value of the target belonging to each category.

[0061] The fully connected module unit, wherein:

[0062] The fully connected acceleration module is used to perform fast fully connected and activation operations on the features output by the ping-pong buffer module. It includes cascaded DSP computing units and activation calculation modules, as well as weight units and bias units connected to the DSP computing units. The DSP computing unit includes N DSP units, which are used to perform parallel multiply-add operations on the features input to the fully connected acceleration module and the values ​​in the weight units and bias units. The values ​​in the weight units and bias units are derived from the weight cache module. The activation calculation module is used to perform ReLU activation calculation on the output of the DSP computing units. Here, N represents the parallelism of the output feature size dimension.

[0063] The DSP computing unit in the fully connected acceleration module can perform N multiplication and addition operations simultaneously, thereby improving system computing efficiency and effectively reducing system latency. At the same time, the size of the DSP computing unit is determined by the parameters of the port of the fully connected acceleration module. By setting this parameter, the size of the DSP computing unit can be adjusted to avoid resource waste caused by the mismatch between the size of the DSP computing unit and the scale of the fully connected operation, thus saving system power consumption. This invention integrates an activation computing module into the fully connected acceleration module. This structure can reduce unnecessary data access between modules, thereby further reducing system latency and power consumption.

[0064] In this embodiment, based on the scale of the fully connected operations in L1, L2 and L3, the fully connected acceleration modules in L1, L2 and L3 are respectively equipped with DSP computing units of scale 8, 2 and 2, where the activation computing modules are all used for ReLU activation computing.

[0065] In this embodiment, each of L1, L2, and L3 contains a ping-pong buffer module. The ping-pong buffer module of L1 is used to buffer the data output by the attention mechanism module in E3, the ping-pong buffer module of L2 is used to buffer the data output by the fully connected acceleration module in L1, and the ping-pong buffer module of L3 is used to buffer the data output by the fully connected acceleration module in L2.

[0066] The weight caching module is used to cache the weight parameters of the fully connected acceleration modules in L1, L2 and L3.

[0067] The AXI interface module includes one AXI4-Lite-Slave interface, one AXI4-Master interface, and corresponding interface control registers. The AXI4-Lite-Slave interface is used to receive control signals and configuration information of the interface control registers sent by the PS, as well as to send status signals of the encoding module unit, decoding module unit, and fully connected module unit. The AXI4-Master interface is used to receive narrowband radar feature data and weight data of the deep feature fusion network sent by the PS, as well as to send calculation results. The interface control registers are used to store the status and function information of the encoding module unit, decoding module unit, and fully connected module unit. In this embodiment, the interface control registers include four 32-bit status registers and 32 64-bit function registers.

Claims

1. A narrowband radar target recognition system based on FPGA and deep feature fusion network, comprising a deep feature fusion network deployed on a hardware platform, characterized in that, The hardware platform employs an FPGA comprising a processor system (PS) and a programmable logic unit (PL); the deep feature fusion network is an accelerator IP core located at the PL, comprising an encoding module unit and cascaded and paralleled decoding and fully connected module units, as well as an AXI interface module located between the encoding module unit and the decoding and fully connected module units; wherein: The encoding module unit includes a cascaded input feature buffer module and multiple encoding modules. Each encoding module includes a cascaded one-dimensional convolution acceleration module, a pooling module, and an attention mechanism module, each with its output end connected to a ping-pong buffer module. The input ends of the one-dimensional convolution acceleration module and the attention mechanism module are also each connected to a weight buffer module. The decoding module unit includes multiple cascaded decoding modules and a reconstruction error module. Each decoding module includes a cascaded upsampling module and a one-dimensional convolution acceleration module, each with its output connected to a ping-pong buffer module. The input of the one-dimensional convolution acceleration module is also connected to a weight cache module. The fully connected module unit includes multiple cascaded fully connected modules. Each fully connected module includes a cascaded ping-pong buffer module and a fully connected acceleration module. The input end of the fully connected acceleration module is also connected to a weight cache module.

2. The system according to claim 1, characterized in that, The processor system terminal (PS) includes a processor and a memory that interacts with it. The processor is used to preprocess narrowband radar feature data received via UART serial port, control the state of accelerator IP cores, postprocess the output results of accelerator IP cores, and output the processing results via UART serial port. The memory is used to cache the weight data of the deep feature fusion network and the processing data of the processor.

3. The system according to claim 1, characterized in that, The encoding module unit, wherein: The input feature caching module is used to cache the narrowband radar feature data input from the AXI interface module; The one-dimensional convolution acceleration module is used to perform fast one-dimensional convolution and activation operations on the features input by the encoding module. It includes a cascaded padding module, a DSP array module, and an activation calculation module, as well as weight and bias units connected to the DSP array module. The padding module is used to zero-padded the edges of the input features. The DSP array module includes... One DSP unit is used to perform parallel multiplication and addition operations on the features output by the padding module and the values ​​in the weight and bias units; the activation calculation module is used to perform ReLU or PReLU activation calculations on the output of the DSP array module; wherein , These represent the parallelism in the output feature size dimension and the parallelism in the output feature channel dimension, respectively. The pooling module is used to perform max pooling calculation on the features output by the ping-pong buffer module connected to its input terminal. The attention mechanism module is used to extract multi-dimensional features from the features output by the ping-pong buffer module cascaded at the output of the pooling module. It includes a cascaded spatial attention mechanism module and a channel attention mechanism module. The spatial attention mechanism includes two pooling modules arranged in parallel and a fully connected acceleration module cascaded at their outputs. The two pooling modules are used for global average pooling and global max pooling calculations, respectively. The channel attention mechanism module includes two pooling modules arranged in parallel and a one-dimensional convolution acceleration module cascaded at their outputs. The two pooling modules are used for global average pooling and global max pooling calculations, respectively. The ping-pong buffer module is used to buffer the data output by the one-dimensional convolution acceleration module, pooling module, or attention mechanism module connected to its input end. The weight cache module is created using BRAM resources within the programmable logic terminal and is used to cache the weight parameters in the accelerator IP core.

4. The system according to claim 3, characterized in that, The size of the DSP array module is determined by the parameters of the port of the one-dimensional convolution acceleration module.

5. The system according to claim 4, characterized in that, The decoding module unit, wherein: The reconstruction error module is used to perform mean square error statistics on the reconstructed feature data output by the decoding module connected to its input terminal and the original features input by the AXI interface module in order to obtain the reconstruction error. The upsampling module is used to upsample the input features of the decoding module.

6. The system according to claim 5, characterized in that, The fully connected module unit, wherein the fully connected acceleration module is used to perform fast fully connected and activation operations on the features output by the ping-pong buffer module, the fully connected acceleration module includes a cascaded DSP computing unit and an activation computing module, as well as a weight unit and a bias unit connected to the DSP computing unit; the DSP computing unit includes One DSP unit is used to perform parallel multiply-accumulate operations on the features input from the fully connected acceleration module and the values ​​in the weight and bias units. An activation calculation module is used to perform ReLU activation calculations on the output of the DSP calculation units. This represents the degree of parallelism in the dimension of the output feature.

7. The system according to claim 6, characterized in that, The size of the DSP computing unit is determined by the parameters of the fully connected acceleration module port.

8. The system according to claim 1, characterized in that, The AXI interface module is used to transmit processing information of the encoding module unit, decoding module unit, and fully connected module unit in the processor and accelerator IP core, and to store the status and function information of the encoding module unit, decoding module unit, and fully connected module unit.

Citation Information

Patent Citations

  • Narrow-band radar aerial target classification method based on depth feature fusion network

    CN117079034A

  • Encoder convenient for LDPC long code implementation on FPGA and encoding method

    CN101489135A

  • Processor, method for operating processor and computer usable medium

    CN108805276A