Sar image target detection and recognition method based on brain-like deep spiking neural network

Abstract

This invention discloses a SAR image target detection and recognition method based on a brain-like deep spiking neural network. It converts SAR images into a pulse event stream format suitable for SNN processing through direct input encoding. An improved membrane potential quantization scheme is employed to reduce quantization errors in the spiking neuron layers. A shallow feature extraction module (LLF-SNN) and a deep feature extraction module (HLF-SNN) are designed to extract multi-level features from the SAR image. Based on this, a Spike-SARYOLO target detection and recognition network model is constructed, including a lightweight SAR image multi-level feature extraction backbone, a cross-scale feature fusion neck network, and a target detection head. Finally, the Spike-SARYOLO network model is trained using a training dataset and inferred using test set data to obtain target detection and recognition results. This invention significantly reduces energy consumption while maintaining high detection accuracy, providing an effective solution for the efficient application of low-cost artificial intelligence technology in the field of SAR image target detection and recognition.

Description

Technical Field

[0001] This invention belongs to the field of radar image processing technology and relates to a SAR image target detection and recognition method based on a brain-like deep spiking neural network. Background Technology

[0002] Synthetic Aperture Radar (SAR) technology has been widely applied in recent years in fields such as ground surveillance, target detection and recognition, disaster monitoring, and military reconnaissance. Compared with optical images, SAR images are unaffected by environmental factors such as weather and lighting, possessing all-weather, all-time, and long-range high-resolution imaging capabilities. Therefore, SAR images exhibit significant advantages in target detection and recognition in complex environments. However, the imaging characteristics of SAR images lead to problems such as inconsistent target scales, complex image backgrounds, and excessive target density, significantly increasing the difficulty of target feature extraction and classification. Furthermore, due to variations in imaging viewpoint, azimuth, and attitude, SAR images of the same target often exhibit significant appearance differences. Therefore, target detection and recognition technologies based on SAR images face numerous challenges.

[0003] Currently, researchers typically employ SAR image target detection and recognition methods based on traditional hand-designed features and those based on deep learning to address SAR image target detection and recognition tasks. Traditional hand-designed feature-based methods rely heavily on manually designed features, making them susceptible to interference from complex backgrounds, resulting in low accuracy, poor processing efficiency, and limited generalization ability. In recent years, with the rapid development of deep learning technology, SAR image target detection and recognition methods based on Artificial Neural Networks (ANNs) have become dominant, especially the YOLO series algorithms, which have achieved high detection accuracy in this field. However, to adapt the network models to more complex recognition scenarios, these deep ANN models often involve massive computational demands, leading to slow operation, insufficient real-time performance, high power consumption, and wasted computing resources, particularly hindering efficient deployment on mobile devices. Therefore, there is an urgent need to develop efficient artificial intelligence algorithms with low computational complexity, low power consumption, and low latency to meet the practical application needs of SAR image target detection and recognition.

[0004] Unlike ANNs, spiking neural networks (SNNs), as a computational model that mimics the working mechanism of biological nervous systems, have shown great potential in fields such as target detection and pattern recognition. SNNs possess powerful spatiotemporal information processing capabilities, enabling the encoding and processing of information through time-series pulses, thus exhibiting higher efficiency and lower energy consumption when handling complex, high-dimensional data. Simultaneously, SNN models possess strong bioinspiration, allowing for more flexible handling of the spatiotemporal transformations of target features in SAR images, effectively avoiding the computational resource waste caused by target sparsity in ANN models during target detection. Therefore, SNNs possess efficient computational capabilities and broad development potential in the field of SAR image target detection and recognition.

[0005] Despite the unique advantages of deep spiking neural networks (SNNs) in SAR image target detection and recognition, they still face a series of technical challenges. When directly training deep SNN models to solve SAR image target detection and recognition tasks, the addition of spiking neurons to the deep SNN network leads to spiking degradation, thus reducing detection performance. Furthermore, the Leaky Integrate-and-Fire (LIF) model inevitably introduces information loss during the quantization of continuous membrane potentials into binary pulses. Therefore, how to efficiently construct deep SNN detection models and optimize membrane potential quantization schemes to develop a SAR image target detection and recognition method based on brain-like deep spiking neural networks, thereby effectively improving SAR image target detection accuracy, reducing computational complexity, and adapting to changes in target features under different scenarios, has become a research topic of significant theoretical and practical value. Summary of the Invention

[0006] The key problem this invention addresses is the high computational cost, slow speed, poor real-time performance, and high power consumption of traditional ANN models in SAR image target detection and recognition tasks. This invention proposes a SAR image target detection and recognition method based on a brain-like deep spiking neural network. The method proposes a Spike-SARYOLO SAR image target detection and recognition algorithm, aiming to reduce the computational complexity and energy consumption of the spiking YOLO network when processing SAR images, thereby improving the performance of the spiking YOLO network in solving SAR image target detection and recognition problems. By designing a reparameterized convolution module and a refined membrane potential quantization scheme, the pulse degradation problem can be effectively alleviated, and the negative impact of quantization errors on model performance can be significantly reduced. This not only improves the model's detection accuracy and robustness but also ensures low power consumption during the inference phase.

[0007] A SAR image target detection and recognition method based on a brain-like deep spiking neural network includes the following steps:

[0008] Step S1: Acquire SAR image data and construct SAR image training, validation, and test sets, specifically as follows:

[0009] After acquiring SAR image data, the original images are first denoised and normalized to improve data quality and model training performance. Then, multiple SAR images are prepared as training, validation, and test sets, with different SAR images included in each dataset. To enhance the diversity of the training set, Mosaic data augmentation techniques are used to simulate different SAR imaging scenarios, thereby increasing the diversity of SAR images.

[0010] Step S2: Spatiotemporally encode the SAR image using a direct input encoding method, transforming the SAR image into an input format that can effectively adapt to the Spike-SARYOLO network model, thereby improving the model's training effect and generalization ability. Specifically:

[0011] The input data dimension of an SNN model is typically represented as: ,in The imaging time step represents the input data. Representative channel, This represents the spatial resolution of the input data. To fully utilize the spatiotemporal capabilities of SNNs, SAR images are spatiotemporally encoded using a direct input encoding method. Assume the original data dimension of a SAR image is... Add a time dimension to it and copy the image. Next, the spatiotemporally encoded input SAR image sequence is obtained, denoted as... Each of them Indicates at the imaging time SAR image data.

[0012] Step S3: Construct the basic information processing unit of the Spike-SARYOLO network model, including the spiking neuron layer. The module consists of two sub-modules: a shallow feature extraction module (LLF-SNN) and a deep feature extraction module (HLF-SNN). Specifically:

[0013] The LIF spiking neuron model achieves a good balance between biological plausibility and computational complexity, and is capable of handling the spatiotemporal information of the input data. In the LIF neuron of this invention, time... membrane potential It combines the time information from the previous moment. and spatial information The state variable, and its dynamic update process calculation expression is:

[0014]

[0015] When membrane potential Exceeding the discharge threshold At this time, LIF neurons fire and generate output pulses. Its amplitude value is determined by the Herveside step function. The decision is made, and its calculation expression is:

[0016]

[0017] In the above formula, if hour ,otherwise If the membrane potential Exceeding the threshold ,pulse When the value is 1, the LIF neuron fires and resets the membrane potential. And subtract This indicates that a pulse signal is emitted. Otherwise, the pulse... When the membrane potential is 0, the LIF neuron continues to accumulate membrane potential, maintains a resting state, and the membrane potential... Through attenuation constant Reduced to :

[0018]

[0019] To reduce the quantization error in the above-mentioned pulse firing process, an Integer Leaky Integrate-and-Fire (ILIF) model will be designed based on the LIF neuron. The pulse firing process of this model can be represented as follows:

[0020]

[0021] In the above formula It is the truncation operator. Indicates will Cut off to ,in This represents the floor function. This is a hyperparameter for determining the maximum firing pulse value in the neuron model. To simplify the representation, the entire ILIF neuron is considered as a single layer of spiking neurons. This layer is trained using integer values ​​during the training phase and converts the integer membrane potential values ​​into binary pulses during inference by extending the virtual time step. Assume that the Spike-SARYOLO network of this invention has a total of... Layer of spiking neurons, where the first The input to the spiking neuron layer during the inference phase is:

[0022]

[0023] In the above formula These are weight parameters. Since matrix multiplication acts as a linear operator, the above equation can be rewritten as:

[0024]

[0025] pulse sequence Since it only contains binary pulse sequences, meaning that multiplication-accumulation operations can be converted into sparse accumulation operations, we can obtain:

[0026]

[0027] Furthermore, the first step of the Spike-SARYOLO network of this invention can be obtained. Layer of spiking neurons The calculation process is as follows:

[0028]

[0029] Based on the ILF spiking neuron layer, and considering the rich semantic knowledge contained in SAR images, this invention designs an LLF-SNN module for extracting shallow features and an HLF-SNN module for extracting deep features from SAR images. The LLF-SNN module consists of a separable convolutional module and a channel convolutional module, and can be represented as follows:

[0030]

[0031] In the formula This is the input data for this module. This indicates the shallow feature extraction module. The separable convolution operation can be further represented as:

[0032]

[0033] in It is a depthwise convolution operation. It is a point convolution operation. This is a channel convolution operation, which can be represented as:

[0034]

[0035] in It is a standard convolution.

[0036] Furthermore, to extract deep features of targets in SAR images, this invention designs an HLF-SNN module, whose channel mixer module employs reparameterized convolution operations. To reduce the number of parameters, the calculation process can be described as follows:

[0037]

[0038]

[0039] in During the inference phase, it reverts to standard convolution. The process of extracting deep features by the HLF-SNN module is as follows:

[0040]

[0041] In summary, the LLF-SNN and HLF-SNN modules have designed different channel convolution modules for application in SAR image target detection and recognition tasks.

[0042] Step S4: Based on the information processing unit defined in step S3, construct the detection and recognition network model Spike-SARYOLO, specifically as follows:

[0043] A lightweight SAR image multi-level feature extraction backbone for Spike-SARYOLO is constructed. This backbone consists of three LLF-SNN modules, three HLF-SNN modules, and one SPPF module established in step S3. Assume the input data at the network input end is... The operation process for Backbone is defined as follows:

[0044]

[0045] in and They represent the first Each feature extraction module , It is a spatial pyramid pooling layer used to convert feature maps of arbitrary size into feature vectors of fixed size. In the design of the SPPF module structure, this invention follows the paradigm of the general YOLO detection algorithm to ensure that the model has good compatibility and efficiency.

[0046] A Spike-SARYOLO SAR image cross-scale feature fusion neck network is constructed. Due to the multi-scale and multi-target characteristics of SAR images, the designed Neck network consists of two parts: a Feature Pyramid Model (FPM) and a Path Aggregation Model (PAM). The FPM module is responsible for transferring features from high-dimensional deep features to low-dimensional shallow features, while the PAM module is responsible for transferring features from low-dimensional shallow features to high-dimensional deep features. By fusing SAR image features of different dimensions, the model's ability to detect and identify multi-scale targets can be effectively improved. The FPM consists of two upsampling modules, one LLF-SNN module, and one HLF-SNN module, while the PAM includes two pulse convolution operations (SConv), one LLF-SNN module, and one HLF-SNN module. The FPM and PAM are directly concatenated through a Concat operation, realizing the fusion of local small-scale features and global large-scale features.

[0047] A Spike-SARYOLO SAR image target detection head is constructed. In the head network designed in this invention, a decoupled head structure is adopted to separate target classification and target regression tasks, achieving more efficient feature learning and task optimization. Simultaneously, an anchor-free detection mechanism is introduced, avoiding the computational complexity caused by traditional anchor box settings, thereby further improving the model's detection accuracy and generalization ability for targets of different shapes in SAR images.

[0048] Furthermore, the Spike-SARYOLO designed in this invention consists of three parts: Backbone, Neck, and Head. Assume the input SAR image sequence at the network input end is... The input data, after feature extraction, fusion, and object detection by the Spike-SARYOLO model, will output the results. It can be represented as:

[0049]

[0050] in This indicates that there are a total of SAR image datasets. Class targets, each Includes prediction categories Category confidence Horizontal coordinates of the center anchor point Vertical coordinates of the center anchor point Anchor frame size .

[0051] Step S5: Construct the MPDIoU loss function and SGD optimizer for the Spike-SARYOLO network model, specifically as follows:

[0052] To better address the problem of detecting and recognizing dense small targets in SAR identification scenarios, an MPDIoU loss function is constructed for the Spike-SARYOLO network model. When the predicted bounding box of a target in a SAR image has the same aspect ratio as the true target bounding box but different dimensions, the traditional IoU (Intersection over Union) metric fails, causing the bounding box regression loss function to be ineffective in optimization, thus limiting the model's convergence speed and accuracy. Therefore, this invention designs an MPDIoU (Minimum Point Distance based IoU) loss function, aiming to further improve the model's classification accuracy and robustness when handling targets in SAR images with uneven distribution, different scales, and diverse shapes. The calculation process is as follows:

[0053]

[0054] In the formula The loss representing the IoU method, The height and width of the SAR image to be detected are... and The calculation process is as follows:

[0055]

[0056]

[0057] in and These are the coordinates of the top-left and bottom-right corners of the predicted bounding box. and These are the coordinates of the top-left and bottom-right corners of the true bounding box. This invention improves the optimization efficiency of bounding box regression by using the coordinates of the top-left and bottom-right corners to determine a unique rectangle and directly minimizing the distance between the predicted bounding box and the true bounding box at these two points.

[0058] When constructing the optimization strategy for the Spike-SARYOLO network model, this invention adopts the Stochastic Gradient Descent (SGD) optimizer, referencing a common deep learning paradigm. The SGD optimizer iteratively updates the weight parameters of the network model to ensure efficient convergence and performance improvement during training.

[0059] Step S6: Input the training set into the Spike-SARYOLO network for training, and iteratively update the weight parameters of the network model through the spatiotemporal backpropagation learning mechanism.

[0060] Step S7: Input the test set into the trained Spike-SARYOLO network for inference and obtain the object detection and recognition results.

[0061] The present invention has the following beneficial effects:

[0062] 1) A Spike-SARYOLO network model was designed to better suit the characteristics of SAR images, which significantly improved the model's detection performance while ensuring low power consumption during the inference phase.

[0063] 2) By avoiding overly complex network structure design, the potential pulse degradation phenomenon in deep spiking neural networks is effectively mitigated, thereby improving the stability and reliability of the model.

[0064] 3) A refined membrane potential quantization scheme was designed to reduce the error generated by spiking neurons during the quantization process, thereby further optimizing the detection accuracy and computational efficiency of the model. Attached Figure Description

[0065] Figure 1 This is a flowchart of the SAR image target detection and recognition method based on a brain-like deep spiking neural network according to the present invention;

[0066] Figure 2 This is a schematic diagram of the SAR image temporal dimension encoding process based on direct input encoding according to the present invention;

[0067] Figure 3 This is a schematic diagram of the composition structure of the shallow feature extraction module LLF-SNN of the present invention;

[0068] Figure 4 This is a schematic diagram of the composition structure of the deep feature SNN extraction module HLF-SNN of the present invention;

[0069] Figure 5 This is a schematic diagram of the composition structure of the backbone for multi-level feature extraction of SAR images using the Spike-SARYOLO network model of the present invention.

[0070] Figure 6 This is a schematic diagram of the composition structure of the Neck network for cross-scale feature fusion of SAR images in the Spike-SARYOLO network model of the present invention.

[0071] Figure 7 This is a schematic diagram of the composition structure of the SAR image target detection head of the Spike-SARYOLO network model of the present invention;

[0072] Figure 8 This is a diagram showing the overall architecture of the Spike-SARYOLO network model of this invention. Detailed Implementation

[0073] The present invention will now be described in more detail with reference to the accompanying drawings and embodiments.

[0074] like Figure 1 As shown, the SAR image target detection and recognition method based on a brain-like deep spiking neural network includes the following steps:

[0075] Step S1: Acquire SAR image data and construct SAR image training, validation, and test sets, specifically as follows:

[0076] After acquiring SAR image data, the original images are first denoised and normalized to improve data quality and model training performance. Then, multiple SAR images are prepared as training, validation, and test sets, with different SAR images included in each dataset. To enhance the diversity of the training set, Mosaic data augmentation techniques are used to simulate different SAR imaging scenarios, thereby increasing the diversity of SAR images.

[0077] Step S2, as follows Figure 2 As shown, spatiotemporal coding of SAR images is performed through direct input coding, transforming the SAR images into an input format that can effectively adapt to the Spike-SARYOLO network model, thereby improving the model's training effect and generalization ability. Specifically:

[0078] The input data dimension of an SNN model is typically represented as: ,in The imaging time step represents the input data. Representative channel, This represents the spatial resolution of the input data. To fully utilize the spatiotemporal capabilities of SNNs, SAR images are spatiotemporally encoded using a direct input encoding method. Assume the original data dimension of a SAR image is... Add a time dimension to it and copy the image. Next, the spatiotemporally encoded input SAR image sequence is obtained, denoted as... Each of them Indicates at the imaging time SAR image data.

[0079] Step S3: Construct the basic information processing unit of the Spike-SARYOLO network model, including the spiking neuron layer. The module consists of two sub-modules: a shallow feature extraction module (LLF-SNN) and a deep feature extraction module (HLF-SNN). Specifically:

[0080] Step S3-1: Construct the spiking neuron layer of the Spike-SARYOLO network model. The LIF spiking neuron model achieves a good balance between biological plausibility and computational complexity, and it can handle the spatiotemporal information of the input data. In the LIF neuron of this invention, time... membrane potential It combines the time information from the previous moment. and spatial information The state variable, and its dynamic update process calculation expression is:

[0081]

[0082] When membrane potential Exceeding the discharge threshold At this time, LIF neurons fire and generate output pulses. Its amplitude value is determined by the Herveside step function. The decision is made, and its calculation expression is:

[0083]

[0084] In the above formula, if hour ,otherwise If the membrane potential Exceeding the threshold ,pulse When the value is 1, the LIF neuron fires and resets the membrane potential. And subtract This indicates that a pulse signal is emitted. Otherwise, the pulse... When the membrane potential is 0, the LIF neuron continues to accumulate membrane potential, maintains a resting state, and the membrane potential... Through attenuation constant Reduced to :

[0085]

[0086] To reduce the quantization error in the above-mentioned pulse firing process, this invention designs an Integer Leaky Integrate-and-Fire (ILIF) model based on LIF neurons. The pulse firing process of this model can be represented as follows:

[0087]

[0088] In the above formula It is the truncation operator. Indicates will Cut off to ,in This represents the floor function. This refers to the hyperparameters that determine the maximum firing pulse value of the neuron model. For simplicity, this invention treats the entire ILIF neuron as a single layer of spiking neurons. This layer is trained using integer values ​​during the training phase and converts the integer membrane potential values ​​into binary pulses during inference by extending the virtual time step. Assume that the Spike-SARYOLO network of this invention has a total of... Layer of spiking neurons, where the first The input to the spiking neuron layer during the inference phase is:

[0089]

[0090] In the above formula These are weight parameters. Since matrix multiplication acts as a linear operator, the above equation can be rewritten as:

[0091]

[0092] pulse sequence Since it only contains binary pulse sequences, meaning that multiplication-accumulation operations can be converted into sparse accumulation operations, we can obtain:

[0093]

[0094] Furthermore, the first step of the Spike-SARYOLO network of this invention can be obtained. Layer of spiking neurons The calculation process is as follows:

[0095]

[0096] Step S3-2, as follows Figure 3 As shown, this invention, based on the ILF spiking neuron layer, and considering the rich semantic knowledge contained in SAR images, designs an LLF-SNN module for extracting shallow features and an HLF-SNN module for extracting deep features from SAR images. The LLF-SNN module consists of a separable convolutional module. A channel convolution module Composition can be represented as:

[0097]

[0098] In the formula This is the input data for this module. This indicates the shallow feature extraction module. The separable convolution operation can be further represented as:

[0099]

[0100] in It is a depthwise convolution operation. It is a point convolution operation. This is a channel convolution operation, which can be represented as:

[0101]

[0102] in It is a standard convolution.

[0103] Furthermore, such as Figure 4 As shown, in order to extract deep features of targets in SAR images, this invention designs an HLF-SNN module, whose channel mixer module employs reparameterized convolution operations. To reduce the number of parameters, the calculation process can be described as follows:

[0104]

[0105]

[0106] in During the inference phase, it reverts to standard convolution. The process of extracting deep features by the HLF-SNN module is as follows:

[0107]

[0108] In summary, the LLF-SNN and HLF-SNN modules have designed different channel convolution modules for application in SAR image target detection and recognition tasks.

[0109] Step S4: Based on the information processing unit defined in step S3, construct the detection and recognition network model Spike-SARYOLO, specifically as follows:

[0110] Step S4-1, as follows Figure 5 As shown, a multi-level feature extraction backbone for Spike-SARYOLO SAR images is constructed. This backbone consists of three LLF-SNN modules, three HLF-SNN modules, and one SPPF module established in step S3. Assume the input data at the network input end is... The operation process for Backbone is defined as follows:

[0111]

[0112] in and They represent the first Each feature extraction module , It is a spatial pyramid pooling layer used to convert feature maps of arbitrary size into feature vectors of fixed size. In the design of the SPPF module structure, this invention follows the paradigm of the general YOLO detection algorithm to ensure that the model has good compatibility and efficiency.

[0113] Step S4-2, as follows Figure 6 As shown, a Spike-SARYOLO SAR image cross-scale feature fusion neck network is constructed. Due to the multi-scale and multi-target characteristics of SAR images, the Neck network designed in this invention consists of two parts: a Feature Pyramid Model (FPM) module and a Path Aggregation Model (PAM) module. The FPM module is responsible for transferring features from high-dimensional deep features to low-dimensional shallow features, while the PAM module is responsible for transferring features from low-dimensional shallow features to high-dimensional deep features. By fusing SAR image features of different dimensions, this invention can effectively improve the model's ability to detect and recognize multi-scale targets. The FPM consists of two upsampling modules, one LLF-SNN module, and one HLF-SNN module, while the PAM includes two pulse convolution operations, one LLF-SNN module, and one HLF-SNN module. The FPM and PAM are directly concatenated through a Concat operation, realizing the fusion of local small-scale features and global large-scale features.

[0114] Step S4-3, as follows Figure 7 As shown, a Spike-SARYOLO SAR image target detection head is constructed. In the head network designed in this invention, a decoupled head structure is adopted to separate the target classification and target regression tasks, thereby achieving more efficient feature learning and task optimization. Simultaneously, this invention introduces an anchor-free detection mechanism, avoiding the computational complexity caused by traditional anchor box settings, thus further improving the model's detection accuracy and generalization ability for targets of different shapes in SAR images.

[0115] Step S4-4, as follows Figure 8 As shown, according to steps S4-1 to S4-3, the Spike-SARYOLO designed in this invention consists of three parts: Backbone, Neck, and Head. Assume the input SAR image sequence at the network input end is... The input data, after feature extraction, fusion, and object detection by the Spike-SARYOLO model, will output the results. It can be represented as:

[0116]

[0117] in This indicates that there are a total of SAR image datasets. Class targets, each Includes prediction categories Category confidence Horizontal coordinates of the center anchor point Vertical coordinates of the center anchor point Anchor frame size .

[0118] Step S5: Construct the MPDIoU loss function and SGD optimizer for the Spike-SARYOLO network model, specifically as follows:

[0119] Step S5-1: To better address the problem of detecting and recognizing dense small targets in SAR identification scenarios, the MPDIoU loss function of the Spike-SARYOLO network model is constructed. When the predicted bounding box of a target in a SAR image has the same aspect ratio as the true target bounding box but different dimensions, the traditional IoU (Intersection over Union) metric fails, causing the bounding box regression loss function to be unable to be effectively optimized, thus limiting the convergence speed and accuracy of the model. Therefore, this invention designs the MPDIoU (Minimum Point Distance based IoU) loss function, aiming to further improve the classification accuracy and robustness of the model when dealing with targets in SAR images with uneven distribution, different scales, and diverse shapes. The calculation process is as follows:

[0120]

[0121] In the formula The loss representing the IoU method, The height and width of the SAR image to be detected are... and The calculation process is as follows:

[0122]

[0123]

[0124] in and These are the coordinates of the top-left and bottom-right corners of the predicted bounding box. and These are the coordinates of the top-left and bottom-right corners of the true bounding box. This invention improves the optimization efficiency of bounding box regression by using the coordinates of the top-left and bottom-right corners to determine a unique rectangle and directly minimizing the distance between the predicted bounding box and the true bounding box at these two points.

[0125] Step S5-2, in constructing the optimization strategy for the Spike-SARYOLO network model, this invention refers to the common deep learning paradigm and adopts the Stochastic Gradient Descent (SGD) optimizer. The SGD optimizer iteratively updates the weight parameters of the network model to ensure efficient convergence and performance improvement during training.

[0126] Step S6: Input the training set into the Spike-SARYOLO network model for training, and iteratively update the weight parameters of the network model through the spatiotemporal backpropagation learning mechanism.

[0127] Step S7: Input the test set into the trained Spike-SARYOLO network model for inference and obtain the object detection and recognition results.

[0128] To verify the effectiveness of the proposed method, target detection experiments were conducted on the publicly available high-resolution SAR target detection and recognition dataset SAR-AIRcraft-1.0. The SAR-AIRcraft-1.0 dataset contains 4386 SAR images with an imaging resolution of 1m and single polarization. This dataset has seven aircraft target categories and is characterized by rich detection and recognition categories, complex imaging scenes, uneven target distribution, strong background noise interference, and multi-scale characteristics. In this invention, the training set, validation set, and test set are divided into SAR image datasets in a 7:1:2 ratio, and the aircraft targets are labeled. This invention uses, but is not limited to, a Windows 10 operating system, an 11th Gen Intel(R) Core(TM) i7-11700 CPU, an NVIDIA GeForce RTX 3060 Ti graphics card with 8GB of video memory, 16GB of RAM, and Python version 3.9. The batch size for training is set to 32, and the training iterations are 600. The trained Spike-SARYOLO network model was tested using the SAR-AIRcraft-1.0 test set. The experimental results show that the method of the present invention achieved a detection accuracy of 88.6%, a recall of 85.9%, an F1 score of 87.2%, and an mAP@75 score of 58.6% on the test set.

[0129] To verify the superiority of the method proposed in this invention, comparative experiments were conducted with several advanced methods in the field of computer vision, including YOLOv3-tiny, YOLOv5, and YOLOv8. Under consistent experimental settings, YOLOv3-tiny achieved a detection accuracy of 70.1%, a recall of 75.9%, an F1 score of 72.9%, and a mAP@75 value of 50.2%; YOLOv5 achieved a detection accuracy of 81.6%, a recall of 78.3%, an F1 score of 79.9%, and a mAP@75 value of 57.7%; and YOLOv8 achieved a detection accuracy of 89.1%, a recall of 87.9%, an F1 score of 88.5%, and a mAP@75 value of 61.9%. The comparative experimental results show that the method of this invention outperforms YOLOv3-tiny and YOLOv5 in terms of detection accuracy, recall, F1 score, and mAP@75 value, indicating that the method of this invention has superior detection performance. Although the detection accuracy, recall, F1 score, and mAP@75 value of the method of this invention are 0.5%, 2%, 1.3%, and 3.3% lower than those of YOLOv8, respectively, its energy consumption in a single inference is significantly lower than that of YOLOv8, with energy consumption of 0.00016 J and 0.009 J, respectively. Specifically, the detection accuracy of the method of this invention is only 0.5% lower than that of YOLOv8, but the energy consumption is reduced by 98.2%. Therefore, the method of this invention can effectively reduce energy consumption while maintaining high detection accuracy, and can be applied to the field of SAR image target detection and recognition.

Claims

1. A SAR image target detection and recognition method based on a brain-like deep spiking neural network, characterized in that, It includes the following steps: Step S1: Acquire SAR image data and construct SAR image training, validation, and test sets, specifically as follows: After acquiring SAR image data, the original images are first denoised and normalized to improve data quality and model training performance. Then, multiple SAR images are prepared as training, validation, and test sets, with different SAR images included in each dataset. To enhance the diversity of the training set, Mosaic data augmentation technology is used to simulate different SAR imaging scenarios, thereby enhancing the diversity of SAR images. Step S2: Spatiotemporally encode the SAR image using a direct input encoding method, transforming the SAR image into an input format effectively adapted to the Spike-SARYOLO network model, thereby improving the model's training performance and generalization ability. Specifically: The input data dimension of an SNN model is typically represented as: ,in The imaging time step represents the input data. Representative channel, This indicates the spatial resolution of the input data; to fully utilize the spatiotemporal capabilities of SNN, spatiotemporal coding of SAR images is performed using a direct input coding method; Assume the original data dimension of a SAR image is 1. Add a time dimension to it and copy the image. Next, the spatiotemporally encoded input SAR image sequence is obtained, denoted as... Each of them Indicates at the imaging time SAR image data on; Step S3: Construct the basic information processing unit of the Spike-SARYOLO network model, including the spiking neuron layer. The module consists of two sub-modules: a shallow feature extraction module (LLF-SNN) and a deep feature extraction module (HLF-SNN). Specifically: The LIF spiking neuron model achieves a good balance between biological plausibility and computational complexity, and is capable of handling the spatiotemporal information of the input data; in the LIF neuron of this invention, time... membrane potential It combines the time information from the previous moment. and spatial information The calculation expression for the state variables in the dynamic update process is: When membrane potential Exceeding the discharge threshold At this time, LIF neurons fire and generate output pulses. Its amplitude value is determined by the Herveside step function. The decision is made, and its calculation expression is: In the formula, if hour ,otherwise If the membrane potential Exceeding the threshold ,pulse When the value is 1, the LIF neuron fires and resets the membrane potential. And subtract This indicates that a pulse signal is being emitted; Otherwise, pulse When the membrane potential is 0, the LIF neuron continues to accumulate membrane potential, maintains a resting state, and the membrane potential... Through attenuation constant Reduced to : To reduce the quantization error in the above-mentioned pulse firing process, an Integer Leaky Integrate-and-Fire (ILIF) model will be designed based on the LIF neuron. The pulse firing process of this model can be represented as follows: In the formula It is the truncation operator. Indicates will Cut off to ,in This represents the floor function. The hyperparameter for the maximum firing pulse value of the neuron model; to simplify the expression, the entire ILIF neuron is considered as a single layer of spiking neurons. This layer is trained with integer values ​​during the training phase and converts integer membrane potential values ​​into binary pulses during inference by extending the virtual time step. Assume the Spike-SARYOLO network has a total of Layer of spiking neurons, where the first The input to the spiking neuron layer during the inference phase is: In the formula These are weight parameters. Since matrix multiplication acts as a linear operator, the above equation can be rewritten as: pulse sequence Since it only contains binary pulse sequences, meaning that multiplication-accumulation operations can be converted into sparse accumulation operations, we can obtain: Furthermore, the first step of obtaining the Spike-SARYOLO network was... Layer of spiking neurons The calculation process is as follows: Based on the ILF spiking neuron layer, and considering the rich semantic knowledge contained in SAR images, an LLF-SNN module for extracting shallow features and an HLF-SNN module for extracting deep features from SAR images were designed. The LLF-SNN module consists of a separable convolutional module and a channel convolutional module, represented as follows: In the formula This is the input data for this module. This indicates the shallow feature extraction module. The separable convolution operation can be further represented as: in It is a depthwise convolution operation. It is a point convolution operation; This is a channel convolution operation, which can be represented as: in It is a standard convolution; Furthermore, to extract deep features of targets in SAR images, an HLF-SNN module was designed. The channel mixer module of the HLF-SNN module employs reparameterized convolution operations. To reduce the number of parameters, the calculation process is described as follows: in During the inference phase, it reverts to standard convolution. The process of extracting deep features by the HLF-SNN module is as follows: In summary, the LLF-SNN and HLF-SNN modules are designed with different channel convolution modules for application in SAR image target detection and recognition tasks. Step S4: Based on the information processing unit defined in step S3, construct the detection and recognition network model Spike-SARYOLO, specifically as follows: A lightweight SAR image multi-level feature extraction backbone, Spike-SARYOLO, is constructed. The backbone consists of three LLF-SNN modules, three HLF-SNN modules, and one SPPF module established in step S3. The input data at the network input end is assumed to be... The operation process for Backbone is defined as follows: in and They represent the first Each feature extraction module , It is a spatial pyramid pooling layer used to convert feature maps of arbitrary size into feature vectors of fixed size. The SPPF module structure design follows the paradigm of the general YOLO detection algorithm to ensure that the model has good compatibility and efficiency. A Spike-SARYOLO SAR image cross-scale feature fusion neck network, Neck, is constructed. Due to the multi-scale and multi-target characteristics of SAR images, the designed Neck network consists of two parts: a Feature Pyramid Model (FPM) and a Path Aggregation Model (PAM). The FPM module is responsible for transferring features from high-dimensional deep features to low-dimensional shallow features, while the PAM module is responsible for transferring features from low-dimensional shallow features to high-dimensional deep features. By fusing SAR image features of different dimensions, the model's ability to detect and recognize multi-scale targets can be effectively improved. The FPM consists of two upsampling modules, one LLF-SNN module, and one HLF-SNN module, while the PAM includes two pulse convolution operations (SConv), one LLF-SNN module, and one HLF-SNN module. The FPM and PAM are directly concatenated through a Concat operation, realizing the fusion of local small-scale features and global large-scale features. A Spike-SARYOLO SAR image target detection head is constructed. In the designed head network, a decoupled head structure is adopted to separate the target classification and target regression tasks, so as to achieve more efficient feature learning and task optimization. At the same time, an anchor-free detection mechanism is introduced to avoid the computational complexity caused by traditional anchor box setting, thereby further improving the model's detection accuracy and generalization ability for targets of different shapes in SAR images. Furthermore, the designed Spike-SARYOLO consists of three parts: Backbone, Neck, and Head; assuming the input SAR image sequence at the network input end is... The input data, after feature extraction, fusion, and object detection by the Spike-SARYOLO model, will output the results. It can be represented as: in This indicates that there are a total of SAR image datasets. Class targets, each Includes prediction categories Category confidence Horizontal coordinates of the center anchor point Vertical coordinates of the center anchor point Anchor frame size ; Step S5: Construct the MPDIoU loss function and SGD optimizer for the Spike-SARYOLO network model, specifically as follows: To better address the problem of detecting and recognizing dense small targets in SAR recognition scenarios, an MPDIoU loss function is constructed for the Spike-SARYOLO network model. When the predicted bounding box of a target in a SAR image has the same aspect ratio but different dimensions from the actual target bounding box, the traditional IoU (Intersection over Union) metric fails, causing the bounding box regression loss function to be ineffective in optimization, thus limiting the model's convergence speed and accuracy. Therefore, an MPDIoU (Minimum Point Distance based IoU) loss function is designed to further improve the model's classification accuracy and robustness when handling targets in SAR images with uneven distribution, different scales, and diverse shapes. The calculation process is as follows: In the formula The loss representing the IoU method, The height and width of the SAR image to be detected are... and The calculation process is as follows: in and These are the coordinates of the top-left and bottom-right corners of the predicted bounding box. and These are the coordinates of the top-left and bottom-right corners of the true bounding box. By using the coordinates of the top-left and bottom-right corners to determine a unique rectangle, and by directly minimizing the distance between the predicted bounding box and the true bounding box at these two points, the optimization efficiency of bounding box regression is improved. When constructing the optimization strategy for the Spike-SARYOLO network model, the Stochastic Gradient Descent (SGD) optimizer is adopted, referring to the general deep learning paradigm. The weight parameters of the network model are iteratively updated through the SGD optimizer to ensure efficient convergence and performance improvement of the model during training. Step S6: Input the training set into the Spike-SARYOLO network for training, and iteratively update the weight parameters of the network model through the spatiotemporal backpropagation learning mechanism; Step S7: Input the test set into the trained Spike-SARYOLO network for inference and obtain the object detection and recognition results.

Images