Lightweight end-to-end sea surface small target detection method and system based on original digital baseband echo signal

Abstract

The invention discloses a lightweight end-to-end sea surface small target detection method and system based on an original digital baseband echo signal, and belongs to the technical field of radar signal processing and feature recognition. The method comprises the following steps: acquiring and preprocessing an original digital baseband echo signal and carrying out framing and slicing; extracting and fusing time domain features in parallel by adopting multi-scale convolution; through time sequence attention modeling, target related time slice response is enhanced, and sea clutter interference is inhibited; and outputting a detection score, adaptively determining a detection threshold according to an expected false alarm rate, and completing target / clutter judgment. According to the method, end-to-end modeling is directly carried out on original echoes, and multi-scale feature extraction and threshold control of time sequence attention and false alarm rate constraint are combined, so that low false alarm, robust detection and lightweight real-time deployment of small sea surface targets under complex sea conditions are realized.

Description

Technical Field

[0001] This invention relates to the field of radar signal processing and feature recognition technology, and more specifically to a lightweight end-to-end method and system for detecting small sea targets based on raw digital baseband echo signals. More particularly, it relates to a lightweight end-to-end method and system for detecting small sea targets using raw digital baseband echo signals as input. Background Technology

[0002] Small target detection technology is mainly used in maritime law enforcement and civilian navigation safety. Its core task is to identify small targets with weak echo characteristics from complex sea clutter backgrounds. Traditional sea surface detection systems are generally based on radar signal processing frameworks. They transmit high-frequency electromagnetic waves and receive sea surface echo signals, and analyze the signal amplitude, phase, and Doppler information to determine whether there are targets on the sea surface.

[0003] Early methods for detecting small targets on the sea surface often employed constant false alarm rate (CFAR) detection and statistical feature modeling strategies. Typical methods included detectors based on amplitude and Doppler statistical characteristics, detectors based on time-frequency analysis characteristics, and detectors that used phase characteristics such as phase difference for discrimination. These methods typically involved estimating parameters of background clutter and setting a detection threshold, within a preset false alarm probability (the false alarm rate depends on the actual needs and can be set independently, generally around 10). -3 Under constraints, abnormal echoes are judged to distinguish between targets and clutter, thereby achieving target detection with a controllable false alarm rate to a certain extent.

[0004] With the development of signal processing and machine learning technologies, researchers have begun to introduce learning algorithms into feature fusion and discrimination processes to construct multi-domain feature spaces for target detection. For example, the K-nearest neighbor-based algorithm introduces complementary information from different feature dimensions to comprehensively determine the detection score from multiple perspectives, such as amplitude, frequency, phase, and temporal changes, making the detection results more robust.

[0005] In recent years, the introduction of deep learning technology has propelled research on sea surface target detection into a new stage. Convolutional neural networks (CNNs) have been widely used for radar echo feature extraction, and their automatic learning capabilities effectively reduce the reliance on manual feature design. Typical examples include detection networks based on residual structures and those incorporating attention mechanisms. These methods achieve high detection accuracy by extracting local and global features from time-frequency maps. In further research, some scholars have proposed dual-channel CNN structures that jointly model and fuse the amplitude and time-frequency information of radar echoes to enhance the model's sensitivity to target signals.

[0006] Overall, existing technologies for detecting small targets on the sea surface have formed a multi-level technical system from traditional statistical detection to intelligent feature learning, covering various methods such as signal-level threshold judgment, feature space classification, and end-to-end detection based on deep learning.

[0007] Although the above methods have achieved certain application results in different scenarios, they still have the following technical limitations under complex sea conditions and actual deployment conditions: 1. Methods based on CFAR and statistical feature modeling typically rely on fixed distribution assumptions (such as Gaussian or Rayleigh distributions), making it difficult to accurately model the non-stationarity and non-Gaussian properties of real sea clutter. In strong clutter scenarios, detection thresholds are prone to failure, leading to a significant increase in false alarm rates.

[0008] 2. Although traditional machine learning methods that fuse multiple features can introduce multi-dimensional feature discrimination information, they still rely on human experience for feature extraction, have poor adaptability to high-noise environments, are difficult to accurately model the time-domain dynamic characteristics of signals, and lack robustness to nonlinear interference.

[0009] 3. Although deep learning models have the ability to learn features automatically, many existing solutions rely on converting the original digital baseband echo signal into a time-frequency image before processing, which may introduce information loss. At the same time, image-based networks usually have a large number of parameters and high inference computation, making it difficult to achieve low-latency and low-power real-time deployment on shipboard or embedded platforms.

[0010] In summary, existing technologies still fall short in terms of robustness, low false alarm rate detection capability, and lightweight deployment in the context of sea clutter interference, making it difficult to fully meet the practical and engineering requirements of next-generation small target detection systems for the sea surface. Therefore, there is an urgent need to develop a lightweight end-to-end detection system that can directly operate on raw digital baseband echo signals and possesses multi-scale feature extraction and temporal modeling capabilities to improve the overall accuracy and deployment efficiency of weak target detection. Summary of the Invention

[0011] In view of this, the present invention provides a lightweight end-to-end method and system for detecting small targets on the sea surface based on raw digital baseband echo signals. The method and system can directly operate on the raw digital baseband echo signals and possess multi-scale feature extraction and temporal modeling capabilities. It aims to solve the problems of insufficient detection performance, high false alarm and false negative rates, high model computational complexity making real-time deployment on shipborne or embedded platforms, and loss of key information due to the conversion of raw signals into time-frequency images for preprocessing in complex sea conditions and strong sea clutter. This improves the overall accuracy, environmental robustness, real-time processing capabilities, and system engineering applicability of small target detection on the sea surface under low signal-to-noise ratio and sparse sample conditions. The "end-to-end" aspect of this invention refers to using the raw digital baseband echo signal as input, undergoing signal preprocessing, multi-scale feature construction, temporal attention modeling, and discrimination and false alarm control, directly outputting a detection score and providing clutter or target decisions based on detection thresholds, thus realizing an integrated detection process from input to output.

[0012] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a lightweight end-to-end method for detecting small targets on the sea surface based on raw digital baseband echo signals, comprising the following steps: Step S100: Acquire the raw digital baseband echo signal received and digitized by the radar system in sea surface monitoring mode; Step S200: Perform signal preprocessing and framing on the original digital baseband echo signal to obtain the preprocessed signal tensor; Step S300: Input the preprocessed signal tensor into the multi-scale feature construction module, extract multi-scale features through parallel convolution branches containing different temporal receptive fields and fuse them to obtain a multi-scale joint feature tensor; Step S400: Input the multi-scale joint feature tensor into the temporal attention modeling module, calculate the temporal correlation and generate attention weights, enhance and suppress the features to obtain the enhanced feature representation; Step S500: Input the enhanced feature representation into the discrimination and false alarm control module, output the detection score, and adaptively determine the detection threshold based on the preset expected false alarm rate (PFA) and background data statistics. Based on the comparison between the detection score and the detection threshold, output the decision result of the target or clutter.

[0013] In a specific implementation scheme, the signal preprocessing and framing in step S200 includes: sequentially performing time-domain filtering, DC component removal, window function processing, and amplitude normalization processing on the original digital baseband echo signal, and framing and slicing according to the preset frame length and frame shift to obtain the preprocessed signal tensor.

[0014] In one specific implementation, the multi-scale feature construction module in step S300 includes at least two parallel convolutional branches, each of which uses a convolutional kernel of different length to perform convolution operations on the preprocessed signal tensor to extract features at different time scales.

[0015] In a specific implementation, the parallel convolutional branch of the multi-scale feature construction module in step S300 includes three branches, which respectively adopt two-dimensional convolution with kernel sizes of (2×3), (2×5) and (2×7).

[0016] In a specific implementation, the temporal attention modeling module in step S400 employs a multi-head attention mechanism, which maps the multi-scale joint feature tensor to query matrices through linear transformation. Key matrix Sum matrix Calculate the query matrix With the key matrix The correlation is calculated and then normalized using Softmax to obtain attention weights. These attention weights are then used to adjust the value matrix. Perform a weighted summation.

[0017] In one specific implementation, the number of attention heads in the multi-head attention mechanism is 8.

[0018] In one specific implementation scheme, the detection score in step S500 is the clutter prediction probability value output by the discrimination and false alarm control module; the adaptive determination of the detection threshold The method is as follows: statistically analyze the empirical distribution of clutter prediction probability values ​​on background data containing only clutter, and select those that meet the conditions. The quantiles are used as the detection threshold. ,in The detection score is... The desired false alarm rate is denoted as .

[0019] Secondly, the present invention provides a lightweight end-to-end small sea surface target detection system based on raw digital baseband echo signals. The system is used to implement the aforementioned lightweight end-to-end small sea surface target detection method based on raw digital baseband echo signals, comprising: The echo signal acquisition module is used to acquire the original digital baseband echo signal received and digitized by the radar system in the sea surface monitoring mode; The signal preprocessing module is used to preprocess and frame the original digital baseband echo signal to obtain a preprocessed signal tensor. A multi-scale feature construction module is used to extract and fuse features from the preprocessed signal tensor through parallel convolutional branches containing different temporal receptive fields to obtain a multi-scale joint feature tensor. The temporal attention modeling module is used to calculate the temporal correlation of the multi-scale joint feature tensor and generate attention weights to enhance and suppress the features, thereby obtaining an enhanced feature representation. The discrimination and false alarm control module is used to process the enhanced feature representation to output a detection score, and adaptively determine the detection threshold based on the preset expected false alarm rate (PFA) and background data statistics. Based on the comparison between the detection score and the detection threshold, the module outputs a decision result of the target or clutter.

[0020] In one specific implementation, the multi-scale feature construction module includes at least two parallel convolutional branches, each of which uses a convolutional kernel of different length to perform convolution operations on the preprocessed signal tensor.

[0021] In one specific implementation, the temporal attention modeling module uses a multi-head attention mechanism to calculate temporal correlation and feature weighting.

[0022] Compared with existing technologies, the lightweight end-to-end method and system for detecting small sea targets based on raw digital baseband echo signals described in this invention are used for automatic detection of small sea targets with high confidence, high sensitivity, and low false alarm rate under complex sea conditions. By constructing an end-to-end modeling system with raw digital baseband echo signals as input, combined with a lightweight multi-scale feature extraction network, a temporal attention modeling module, and an adaptive threshold control algorithm based on the expected false alarm rate, it achieves effective differentiation and robust decision-making between weak target signals and complex sea clutter interference, effectively improving the detection sensitivity, environmental adaptability, and real-time processing capability of the detection system under low signal-to-clutter conditions, and has the following beneficial effects: 1. This invention employs a technique of end-to-end modeling directly on the original digital baseband echo signal, avoiding information loss caused by converting the signal into a time-frequency image. It also combines a multi-scale convolutional structure to extract signal features at different time scales in parallel and fuse them, enabling the detection system to make fuller use of the temporal details and contextual information of the signal, thereby improving the sensitivity of detection.

[0023] 2. This invention introduces a temporal attention module to model the correlation of features in the time dimension, dynamically allocates attention weights to highlight relevant segments of the target and suppress transient interference, and adopts an adaptive threshold control strategy that combines real-time background noise statistics for target discrimination. This enables the detection system to adapt to the non-Gaussianity and time-varying characteristics of sea clutter and maintain the stability of detection performance when facing different signal-to-clutter ratios and severe sea conditions.

[0024] 3. In the process of constructing the feature extraction and temporal modeling network, this invention adopts a compact combination of convolutional coding and attention modeling, and reduces the number of model parameters and inference computation overhead by controlling network depth, channel size and computation path. This enables the detection system to maintain detection performance while having a smaller model size and lower deployment resource consumption, thus making it suitable for application scenarios with strict requirements for real-time performance and resource constraints, such as shipborne platforms and unmanned surface vessel embedded systems. This achieves the integration from algorithm to engineering application. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0026] Figure 1 This is a flowchart of a lightweight end-to-end small target detection method for the sea surface based on the original digital baseband echo signal, as described in this invention.

[0027] Figure 2 This is a schematic diagram of the echo signal acquisition module of the present invention.

[0028] Figure 3 This is a schematic diagram of the signal preprocessing module of the present invention.

[0029] Figure 4 This is a schematic diagram of the multi-scale feature construction module of the present invention.

[0030] Figure 5 This is a schematic diagram of the temporal attention modeling module of the present invention.

[0031] Figure 6 This is a flowchart illustrating the detection and false alarm control module of the present invention. Detailed Implementation

[0032] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] This invention describes a lightweight end-to-end small target detection method for the sea surface based on raw digital baseband echo signals. Applied to small target identification on the sea surface, it constructs an end-to-end framework that directly processes the raw digital baseband echo signals. This framework sequentially includes: echo signal acquisition, signal preprocessing, feature extraction based on a lightweight multi-scale convolutional network, temporal attention modeling, and adaptive threshold discrimination based on the Probability of False Alarm (PFA). Furthermore, during model training, a weighted loss function is used to address the sample imbalance problem and improve model robustness.

[0034] In one embodiment, the weighted loss function can take the form of weighted binary cross-entropy, and for a single training sample, its loss can be expressed as: in, This represents the true label of the sample (target is 1, clutter is 0). This represents the target prediction probability output by the network. For target class weights, These are clutter weights. The weights can be determined using a static preset method: counting the number of target samples in the training dataset. With clutter sample number And set weights according to the inverse relationship between category frequency, for example: In another alternative embodiment, the weights can also be updated in segments according to training epochs: every certain number of epochs, the proportion of classes in the training subset / batch is recalculated and updated. This is to adapt to fluctuations in category distribution under different sea conditions or different data batches. During training, the network forwards the preprocessed signal tensor of the input to predict the output probability. Then calculate the loss based on the actual label. The parameters of the multi-scale convolutional network and the temporal attention module are updated through backpropagation, thereby embedding the loss function into the end-to-end training framework.

[0035] It should be noted that the echo signal acquisition, data acquisition, scene recording, and other activities involved in this invention are all carried out in compliance with relevant regulations and with the authorization to use the corresponding equipment.

[0036] In this specification, to avoid ambiguity in terminology, the relevant objects are uniformly defined as follows: The sea surface scattered echo received by the radar antenna is amplified, down-converted, and filtered in the receiving link, and then converted from analog to digital to obtain the original digital baseband echo signal. The original digital baseband echo signal is the baseband echo sequence digitized at the receiving end, containing two channels of information (I / Q dual channels) of in-phase and quadrature components, which can be equivalently represented as a complex baseband echo sequence or its two-dimensional real-valued representation. After preprocessing and framing in step S200, the original digital baseband echo signal forms a preprocessed signal tensor. Step S300 yields a multi-scale joint feature tensor, and step S400 yields an enhanced feature representation. The detection score output in step S500 is the clutter prediction probability value among the two types of prediction probabilities obtained by the Softmax function, with a value range of [0,1]. It is used to compare with the detection threshold to output "clutter decision" or "target decision". To ensure symbol consistency, the "detection score" is uniformly denoted as... The "detection threshold" is uniformly recorded as The expected false alarm rate is uniformly denoted as .

[0037] The preprocessed signal tensor output in step S200 can be represented as follows during single-frame inference: The first dimension "2" corresponds to the I / Q dual channel, and the second dimension... The frame length is specified; during batch training or batch inference, multiple frame samples can be stacked. ,in This is the batch size. Unless otherwise specified, subsequent formulas are in single-frame format. The batch format can be computed independently and in parallel according to the sample dimension.

[0038] The following provides a detailed description of each component and workflow of the framework.

[0039] 1. Composition and Module Functions of the Detection System The detection system of the present invention mainly includes: an echo signal acquisition module, a signal preprocessing module, a multi-scale feature construction module, a temporal attention modeling module, a discrimination and false alarm control module, and a system deployment and operation module.

[0040] The echo signal acquisition module is used to collect and digitize the sea surface scattered echoes. It can be composed of a radar antenna and RF front-end, low-noise amplifier, down-conversion and filtering link, analog-to-digital converter, etc., and is used to output the raw digital baseband echo signal to provide input for subsequent processing.

[0041] The signal preprocessing module performs filtering, DC removal, window function processing, and normalization on the original digital baseband echo signal, and performs frame slicing according to the preset frame structure to form a preprocessed signal tensor that meets the network input format, thereby improving the stability and consistency of the input data.

[0042] The multi-scale feature construction module is used to extract multi-scale feature representations with different temporal receptive fields from the preprocessed time-series signal. This module extracts short-term transient, local context, and long-term dependency features through a parallel convolutional branch structure. After concatenation, it combines pooling statistics and one-dimensional convolution to generate weight coefficients, achieving adaptive enhancement and suppression of multi-scale features. Finally, it fuses the features at each scale into a unified multi-scale joint feature tensor, providing rich representational information for subsequent modeling.

[0043] The temporal attention modeling module takes a multi-scale joint feature tensor as input, establishes temporal dependencies and generates attention weights to enhance key time segments related to the target echo, and suppresses random fluctuations or transient spikes in the sea clutter background, thereby outputting an enhanced feature representation for subsequent decision-making.

[0044] The discrimination and false alarm control module is used to output detection results and detection scores. This module compresses and maps the enhanced feature representation, outputting a detection score for threshold comparison; wherein, the detection score is the clutter prediction probability value. Based on the preset expected false alarm rate (PFA), the detection threshold is adaptively determined through background statistics. Under the premise of satisfying the false alarm constraint, it outputs either a "clutter decision" or a "target decision," along with the corresponding detection score and associated index information.

[0045] The system deployment and operation module is used to deploy algorithms on the computer, communicate via interfaces, and run online inference. It provides external control and parameter configuration interfaces, supports operation on general-purpose processors, graphics processors, programmable logic devices, or embedded platforms, and enables switching between real-time detection and offline analysis modes. It can also output detection results, detection scores, and related index information to a host computer or task system.

[0046] The overall working principle of the detection system is as follows: the radar front end collects echo signals, which are then processed by the preprocessing module to generate standardized input tensors. These tensors are then input to the feature extraction and temporal modeling module to complete feature learning and target enhancement. Subsequently, the discrimination and false alarm control module outputs the target detection results, detection scores, and time frame indexes, thereby achieving automatic detection and recognition of small targets on the sea surface.

[0047] The following combination Figure 1 The flowchart shown illustrates the execution steps of this method in detail with a specific embodiment.

[0048] In one embodiment, for ease of reproduction and implementation, the preprocessing and inference can use the following parameter combination: frame length A value of 1024 can be used, and the frame shift can be 200 (adjacent frames overlap to improve the continuity of weak targets and the statistical stability of the threshold). A Hamming window can be used as the window function to reduce abrupt changes in frame boundaries; the length of the multi-scale convolutional branch kernel... This is used to cover three types of time-scale information: transient spikes, local textures, and longer trends; the number of temporal attention heads is set to 8 to balance modeling capability and computational cost; PFA is the default value. The above parameters are not unique values. Those skilled in the art can adjust them based on radar sampling rate, dwell time, length of sea clutter stable interval, and platform computing power constraints. The frame length, among other things... The frame shift can be selected within the range of 512 to 4096 to balance "temporal context coverage" and "non-stationarity introduction". ~ The range of parameters is selected to balance real-time performance and statistical stability, and the number of attention heads can be selected from 4 to 8 to balance modeling capability and inference complexity. By setting the above parameters, the stability of subsequent threshold calibration based on clutter statistics can be improved, thereby supporting low false alarm control.

[0049] In this embodiment, the radar operates in sea surface surveillance mode, and sea surface echo signals are acquired through an antenna array, a down-conversion module, and an analog-to-digital conversion module to obtain raw digital baseband echo signals that include sea clutter, sea spikes, sea surface textures, and weak target echoes.

[0050] The acquired raw digital baseband echo signal is first input to the signal preprocessing module, where it undergoes bandpass filtering, DC bias removal, normalization, and framing operations to ensure that the subsequent neural network input has a stable amplitude scale and a fixed length format.

[0051] The preprocessed signal tensor is used by the multi-scale feature construction module of this invention, with convolutional kernels of different lengths modeling the short-term transient features, mid-range texture information, and longer-term dependent structures of the signal. The feature maps output by each convolutional branch are concatenated and normalized to form a multi-channel feature tensor.

[0052] Subsequently, a temporal attention module is used to calculate the correlation of the signal in the time dimension, dynamically allocate weights, improve the intensity of local region feature representation related to the target, and reduce the influence of transient interference such as spikes and impulse noise, making the entire feature representation more robust.

[0053] The tensor, after feature enhancement, is input to the discrimination and false alarm control module. This module uses a fully connected layer to compress the features and obtain a detection score. Combining the expected false alarm rate with the real-time estimated background noise statistics, the system automatically determines the detection threshold to achieve stable detection of weak targets.

[0054] In this embodiment, the detection system of the present invention is deployed in a shipborne processor and an embedded platform of an unmanned surface vessel. After lightweight design optimization, the inference time of the entire model can meet the requirements of real-time maritime surveillance under typical hardware conditions.

[0055] Figure 1 This is one of the flowcharts of a method for detecting floating targets on the sea surface provided by the present invention, as shown below. Figure 1 The present invention provides a method for detecting floating targets on the sea surface, which can be applied to the field of detecting floating targets on the sea surface and also to the field of detecting small floating targets on the sea surface. If applied to the field of detecting small floating targets on the sea surface, the target of the technical solution is adapted to be a small target. The method provided by the present invention includes steps S100-S600.

[0056] Step S100: Acquire the raw digital baseband echo signal. The radar system operates in sea surface monitoring mode, transmitting electromagnetic waves through the antenna array and receiving the sea surface scattered echoes. Figure 2 A schematic diagram of the echo signal acquisition module is provided. The sea surface scattered echo is amplified by a low-noise amplifier, then frequency-converted by a down-converter, and subsequently filtered by a bandpass filter to remove irrelevant frequency band interference. An analog-to-digital converter then converts the analog echo into a raw digital baseband echo signal containing sea clutter, sea spikes, and echoes from potential weak targets. The clutter provided in this embodiment refers to interference signals or noise other than the signal reflected back from the target object during radar detection.

[0057] Step S200: Signal preprocessing and framing to obtain the preprocessed signal tensor. Figure 3A schematic diagram of the signal preprocessing structure is provided, where the preprocessing flow includes filtering, DC component removal, window function processing, normalization, and frame slicing. The raw digital baseband echo signal obtained in step S100 is input to the signal preprocessing module. The signal preprocessing module sequentially performs time-domain filtering, DC component removal, window function processing, and amplitude normalization on the raw digital baseband echo signal. The time-domain filtering is a digital baseband filtering method used to suppress residual narrowband / out-of-band interference after analog-to-digital conversion, which differs from the front-end anti-aliasing filtering. DC component removal is used to eliminate baseline offset introduced by the receiving link. Window function processing is used to reduce spectral leakage and abrupt changes at frame boundaries. Amplitude normalization is used to align the amplitude scales of different batches of echo data. Subsequently, the signal preprocessing module slices the continuous echoes according to a preset frame length and frame shift to obtain fixed-length signal segments, and organizes these signal segments into a preprocessed signal tensor output. Through the above processing, the preprocessed signal has a uniform amplitude scale and a clear temporal organization, facilitating stable feature learning and discrimination by subsequent deep learning detection models.

[0058] Step S300: As Figure 4 As shown, the preprocessed signal tensor output from step S200 is input into a multi-scale feature construction module. This module uses a parallel multi-branch structure of a convolutional neural network to extract features from the signal. The multi-branch structure consists of three parallel convolutional branches, each using a convolutional kernel of different lengths to form different temporal receptive fields. These are used to extract transient change features, local context features, and long-term correlation trend features from the same input. The first layer of each branch uses a two-dimensional convolution with a kernel size of (2×k), where 2 is used to cover the I / Q channels to achieve cross-channel fusion. k takes values ​​of 3, 5, and 7, forming three branches: a two-dimensional convolution (2×3), a two-dimensional convolution (2×5), and a two-dimensional convolution (2×7), corresponding to the extraction of transient spike features, mesoscale features, and long-term trend features. In one embodiment, each signal frame... Where "2" corresponds to I / Q dual channels, " is the frame length, which can be 1024 in one embodiment; the two-dimensional convolution forms a length of " in the time dimension The receptive field is essentially used for convolutional modeling of temporal echoes, while cross-channel fusion is achieved by covering the I / Q with the first dimension.

[0059] The corresponding convolution and nonlinear mapping process can be represented as: in, This represents the convolution operation; and These are the kernel parameters and bias parameters for the branch with kernel length k, respectively; This indicates a batch normalization operation; This represents a nonlinear activation function, and in this embodiment... For ReLU; This represents the preprocessed signal tensor output in step S200; This represents the convolutional output feature map of the branch with kernel length k. Then, the outputs of each branch are concatenated along the channel dimension to obtain a multi-scale feature tensor: , in This indicates that splicing is performed in the channel dimension. In one embodiment, each branch maintains a consistent time dimension length by using the same time length and appropriate padding, thereby allowing splicing to be performed in the channel dimension. Furthermore, the multi-scale feature construction module... Max pooling and average pooling are performed separately to form statistical descriptions. These statistical descriptions are then sequentially input into a one-dimensional convolution, ReLU activation, and another one-dimensional convolution to generate two sets of weights. The two sets of weights obtained from the max pooling branch and the average pooling branch are summed element-wise to form the final weight coefficients. These weight coefficients are then compared with... By performing channel-by-channel multiplication, adaptive enhancement and suppression of multi-scale features are achieved, resulting in a multi-scale joint feature tensor. and the As input to step S400.

[0060] Step S400: Perform temporal attention modeling, with the module structure as follows: Figure 5 As shown. The multi-scale joint feature tensor output in step S300 is... The input is a temporal attention modeling module, which generates query matrices through linear transformations. Key matrix Sum matrix : in These are trainable linear transformation parameters, used to transform the multi-scale joint feature tensor. Mapping on the channel feature dimension to obtain Subsequently, temporal correlation calculations are performed to obtain attention weights, and softmax normalization is used to obtain weight coefficients, which are then applied to... Feature-weighted summation is performed. The temporal attention modeling module employs a multi-head attention mechanism with 8 attention heads. Soon The feature dimension is divided into 8 attention heads, which are computed in parallel. Each attention head obtains a weighted output in the following form: Subscript Indicates the first One point of attention, The key vector dimension is used to scale the relevance calculation; the Softmax normalization is performed along the key / time index dimension to obtain the attention weights at different times. The outputs of each attention head are concatenated and linearly mapped to obtain the enhanced feature representation. Through the above process, the response in key time segments consistent with the target echo can be highlighted, and the background of sea clutter and non-target transient disturbances can be suppressed, outputting enhanced feature representations for subsequent discrimination. .

[0061] Step S500: Input the output feature from step S400 into the discrimination and false alarm control module. See the flowchart for this module. Figure 6 The module takes the input feature map as input, compresses the features through a global average pooling layer, then passes through a linear layer and a dropout layer to obtain the classification output, and uses a Softmax activation function to obtain the predicted probability of "clutter / target"; further, it extracts the clutter prediction probability value as the detection score. The discrimination and false alarm control module calibrates the detection threshold based on a preset expected false alarm rate (PFA): it statistically calculates the clutter prediction probability on background data containing only clutter. The empirical distribution is selected such that it satisfies the following in the clutter background: quantiles as detection thresholds During operation, when the clutter prediction probability... Greater than It outputs "clutter decision" when the clutter is present, and "target decision" when the target is not present. It also outputs the detection score and associated index information simultaneously, thereby achieving reproducible low false alarm control under different sea conditions and background fluctuations.

[0062] Since this embodiment uses "clutter prediction probability" as the detection score s, a higher score indicates a greater likelihood of clutter. Therefore, the threshold comparison direction is set as follows: when... The system outputs a "clutter decision" if the clutter condition is met, and a "target decision" otherwise. The "background data containing only clutter" can be obtained as follows: During the offline calibration phase, time periods and spatial units in the monitored scene confirmed to be target-free by manual recording or external information can be selected as clutter samples; alternatively, background distance units and azimuth sectors far from the waterway / high-incidence target area can be selected as clutter sampling areas, and abnormally high-energy frames can be removed to reduce the probability of target intrusion. During the online update phase, a clutter buffer can be maintained within the system's sliding time window, and a "detection result removal" strategy can be adopted: frames judged as targets or highly suspected targets (e.g., scores crossing thresholds and having temporal continuity) are excluded from the background buffer, and only the clutter score distribution is updated for the remaining samples to achieve online adaptive updating of the threshold. In one embodiment, the threshold calibration process includes: step a) collecting clutter background data and completing preprocessing and framing consistent with online methods; step b) inputting the clutter frame into the trained model to obtain the clutter prediction probability score s; step c) statistically analyzing the empirical distribution of the score s and calculating the quantiles; step d) selecting the quantiles that satisfy the desired false alarm rate constraint as the detection threshold τ and writing it into the configuration; step e) (optional) periodically updating steps b) to d) during operation using a sliding window, and verifying the stability of the update results before taking effect. In one embodiment, the desired false alarm rate PFA can be taken as... Within the range, to match the acceptable number of false alarms per unit time / unit scan; where, while balancing alarm burden and weak target detection requirements, PFA can be set to the default value. It also allows adjustments to the task configuration based on platform alarm capacity and sea state intensity.

[0063] To verify the quality of clutter background data, a consistency check can be performed on the score distribution within the background buffer. For example, check whether the quantile drift of the score distribution in adjacent time windows exceeds a preset threshold, or check whether the proportion of high-scoring abnormal samples exceeds the upper limit. When the check fails, the threshold update can be frozen and a prompt to re-acquire / select background data can be displayed to avoid target mixing that could cause threshold distortion.

[0064] Step S600: Deploy the trained and parameter-fixed deep learning detection model on a shipboard platform, an unmanned surface vessel embedded processing system, or a shore-based monitoring device. The system deployment and operation module is responsible for model loading, hardware adaptation, calling operation parameters, and monitoring operation status, and interacts with the host computer or monitoring terminal through the communication interface.

[0065] The deployed system receives radar echo signals in real time. Driven by the system deployment and operation module, it performs an online detection loop, cyclically executing steps S100 to S500. The detection decision, detection score, and associated index information output in step S500 are reported to a host computer, monitoring terminal, or automatic alarm system to achieve online monitoring and intelligent identification of small targets on the sea surface. Optionally, model training and threshold calibration are completed offline before deployment. The training data can cover multiple sea state backgrounds and target samples. Default values ​​and detection thresholds can be calibrated based on on-site background data and then written into configuration parameters to meet engineering operation requirements.

[0066] In one embodiment, the model training process includes at least: the organization and annotation method of the training data, the loss function used and its design considerations, and the basic training configuration. Specifically, the training data consists of samples after frame segmentation, each sample containing I / Q dual-channel signal frames and their labels; the labels can be obtained through manual annotation, external observation information, or verification of existing detection records, and are divided into training set / validation set / test set. During training, a mini-batch input method is used (batch size B can be 16-256), the weighted loss function is used as the optimization objective, and a gradient-based optimizer (the Adam optimizer in this method) is used to update the network parameters; the learning rate, training epochs, and other configurations can be set according to the performance of the validation set and the convergence situation, and early stopping or learning rate decay strategies can be used to improve training stability.

[0067] The various embodiments described in this specification are presented in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A lightweight end-to-end method for detecting small sea surface targets based on raw digital baseband echo signals, characterized in that, Includes the following steps: Step S100: Acquire the raw digital baseband echo signal received and digitized by the radar system in sea surface monitoring mode; Step S200: Perform signal preprocessing and framing on the original digital baseband echo signal to obtain the preprocessed signal tensor; Step S300: Input the preprocessed signal tensor into the multi-scale feature construction module, extract multi-scale features through parallel convolution branches containing different temporal receptive fields and fuse them to obtain a multi-scale joint feature tensor; Step S400: Input the multi-scale joint feature tensor into the temporal attention modeling module, calculate the temporal correlation and generate attention weights, enhance and suppress the features to obtain the enhanced feature representation; Step S5 00: Input the enhanced feature representation into the discrimination and false alarm control module, output the detection score, and adaptively determine the detection threshold based on the preset expected false alarm rate (PFA) and background data statistics. Based on the comparison between the detection score and the detection threshold, output the decision result of the target or clutter.

2. The lightweight end-to-end sea surface small target detection method based on raw digital baseband echo signal according to claim 1, characterized in that, The signal preprocessing and framing in step S200 includes: sequentially performing time-domain filtering, DC component removal, window function processing, and amplitude normalization on the original digital baseband echo signal, and framing and slicing according to the preset frame length and frame shift to obtain the preprocessed signal tensor.

3. A lightweight end-to-end sea surface small target detection method based on raw digital baseband echo signals according to claim 1, characterized in that, The multi-scale feature construction module in step S300 includes at least two parallel convolutional branches. Each convolutional branch uses a convolutional kernel of different length to perform convolution operations on the preprocessed signal tensor to extract features at different time scales.

4. A lightweight end-to-end sea surface small target detection method based on raw digital baseband echo signals according to claim 3, characterized in that, The parallel convolution branch of the multi-scale feature construction module in step S300 includes three branches, which respectively adopt two-dimensional convolution with kernel sizes of (2×3), (2×5) and (2×7).

5. A lightweight end-to-end sea surface small target detection method based on raw digital baseband echo signals according to claim 1, characterized in that, The temporal attention modeling module in step S400 employs a multi-head attention mechanism, which maps the multi-scale joint feature tensor to query matrices through linear transformation. Key matrix Sum matrix Calculate the query matrix With the key matrix The correlation is calculated and then normalized using Softmax to obtain attention weights. These attention weights are then used to adjust the value matrix. Perform a weighted summation.

6. A lightweight end-to-end sea surface small target detection method based on raw digital baseband echo signals according to claim 5, characterized in that, The multi-head attention mechanism has 8 attention heads.

7. A lightweight end-to-end sea surface small target detection method based on raw digital baseband echo signals according to claim 1, characterized in that, The detection score in step S500 is the clutter prediction probability value output by the discrimination and false alarm control module; the adaptive determination of the detection threshold The method is as follows: statistically analyze the empirical distribution of clutter prediction probability values ​​on background data containing only clutter, and select those that meet the conditions. The quantiles are used as the detection threshold. ,in The detection score is... The desired false alarm rate is denoted as .

8. A lightweight end-to-end small sea surface target detection system based on raw digital baseband echo signals, characterized in that, The system is used to implement a lightweight end-to-end small sea surface target detection method based on raw digital baseband echo signals, as described in any one of claims 1-7, comprising: The echo signal acquisition module is used to acquire the original digital baseband echo signal received and digitized by the radar system in the sea surface monitoring mode; The signal preprocessing module is used to preprocess and frame the original digital baseband echo signal to obtain a preprocessed signal tensor. A multi-scale feature construction module is used to extract and fuse features from the preprocessed signal tensor through parallel convolutional branches containing different temporal receptive fields to obtain a multi-scale joint feature tensor. The temporal attention modeling module is used to calculate the temporal correlation of the multi-scale joint feature tensor and generate attention weights to enhance and suppress the features, thereby obtaining an enhanced feature representation. The discrimination and false alarm control module is used to process the enhanced feature representation to output a detection score, and adaptively determine the detection threshold based on the preset expected false alarm rate (PFA) and background data statistics. Based on the comparison between the detection score and the detection threshold, the module outputs a decision result of the target or clutter.

9. A lightweight end-to-end small sea surface target detection system based on raw digital baseband echo signals according to claim 8, characterized in that, The multi-scale feature construction module contains at least two parallel convolutional branches, each of which uses a convolutional kernel of different length to perform convolution operations on the preprocessed signal tensor.

10. A lightweight end-to-end small sea surface target detection system based on raw digital baseband echo signals according to claim 8, characterized in that, The temporal attention modeling module uses a multi-head attention mechanism to calculate temporal correlation and feature weighting.

Images