CN-122017777-A - Small target detection method of complex dynamic attention mechanism

CN122017777ACN 122017777 ACN122017777 ACN 122017777ACN-122017777-A

Abstract

The invention discloses a small target detection method of a complex dynamic attention mechanism, which comprises the steps of taking a complex matrix obtained by pulse pressure and moving target detection processing of target reflection echoes received by a radar as input, establishing a background, clutter and target three-classification labeling chart as a label and setting an effective mask by constructing an I/Q (input/output) double-channel standardized tensor as a training sample, constructing a small target detection and segmentation network based on U-Net in a training stage, embedding a complex domain dynamic attention module after a complex residual block of an encoder, wherein the complex domain dynamic attention module comprises a complex domain channel attention module, a multi-scale space attention module and a dynamic fusion gating module which are respectively used for carrying out channel enhancement and space enhancement on complex features, adopting a mixed loss function in training, outputting a prediction class in an inference stage, and obtaining a detection result by a target class mask.

Inventors

SU JIA
ZHANG MANQI
WANG HAITAO
CHEN SHICHAO
WANG LING
FAN YIFEI
GUO ZIXUN
TAO MINGLIANG

Assignees

西北工业大学

Dates

Publication Date: 20260512
Application Date: 20260119

Claims (10)

1. A method for detecting a small target of a complex dynamic attention mechanism, comprising: The method comprises the steps of taking a complex matrix obtained by pulse pressure and moving target detection processing of target reflection echoes received by a radar as input, establishing a background, clutter and target three-classification label graph as labels and setting an effective mask by constructing an I/Q (input/output) dual-channel standardized tensor as a training sample, constructing a small target detection and segmentation network based on U-Net in a training stage, embedding a complex domain dynamic attention module after a complex residual block of an encoder, wherein the complex domain dynamic attention module comprises a complex domain channel attention module, a multi-scale space attention module and a dynamic fusion gating module which are respectively used for carrying out channel enhancement and space enhancement on complex features and then carrying out dynamic fusion, adopting a mixing loss function in training, outputting a prediction category in an inference stage, and obtaining a detection result by a target category mask.
2. The method for detecting the small target of the complex dynamic attention mechanism according to claim 1, wherein the small target detection and segmentation network takes U-Net as a backbone and comprises an encoder, a decoder and an output head, wherein the encoder is formed by stacking a plurality of complex residual blocks, a complex domain dynamic attention module is embedded in the complex residual blocks, the complex residual blocks are represented by an equivalent complex number of I/Q double channels, and input complex features and output complex features of any layer in the network are represented as follows: Wherein, the Representing the layer of complex features; And (3) with Respectively representing real part characteristics and imaginary part characteristics thereof; Representing the number of characteristic channels; a spatial dimension representing a feature of the layer; Representation of Real valued tensor space of dimensions.
3. The small object detection method of complex dynamic attention mechanism of claim 1, wherein the feature extraction process of complex residual blocks in the encoder comprises: Let complex convolution kernel be For inputting complex characteristics Complex convolution is performed with the equivalent real output: ; is the real part characteristic of the convolution output; Is the real/imaginary part characteristic of the input; real/imaginary parameters that are convolution kernels; Representing a two-dimensional convolution operation; The equivalent imaginary output of complex convolution is 。
4. The method for small object detection in a complex dynamic attention mechanism of claim 1, wherein the processing of the complex domain channel attention module comprises: first, outputting real part characteristic and imaginary part characteristic of complex residual block Global average pooling is carried out in space dimension and the channel description vectors are obtained by splicing Subsequent channel description vector Generating channel weights through two-layer full-connection Weighting the channels Splitting into real part channel weights Channel weights with imaginary part Will (i) be And (3) with Respectively acting on real and imaginary features Obtaining real part characteristic and imaginary part characteristic after channel enhancement 。
5. The method for small object detection by complex dynamic attention mechanisms of claim 1, wherein the processing by the spatial attention module comprises: first, for the real part characteristic and the imaginary part characteristic after channel enhancement Computing amplitude feature tensors Then, for amplitude characteristic tensor Carrying out channel dimension average pooling and maximum pooling and splicing to obtain a spliced feature vector Adopting multi-scale convolution branches to aim at splicing characteristic vectors Extracting space information and splicing to obtain multi-scale convolution characteristics Multi-scale convolution features After convolution fusion and through an activation function, a space attention weight graph is obtained The space attention weight map will be Acting on real and imaginary characteristics of the channel enhancement On the basis of the real part characteristic and the imaginary part characteristic after the space enhancement 。
6. The method for small object detection by plural dynamic attention mechanisms according to claim 1, wherein the processing procedure of the dynamic fusion gating module comprises: the real part characteristic and the imaginary part characteristic of the channel after the enhancement are carried out Real and imaginary characteristics after spatial enhancement Stitching in the channel dimension to construct input features Dynamic fusion gating module aims at input characteristics Performing convolution processing, activation function processing and convolution processing in sequence, and performing Softmax normalization on the channel attention branch and the space attention branch respectively to obtain a channel attention branch weight map Space attention branching weight map ; By means of And Real and imaginary characteristics of enhanced channel Real and imaginary characteristics after spatial enhancement Dynamic fusion is carried out to obtain the real part characteristic and the imaginary part characteristic of the output of the complex domain dynamic attention module 。
7. The method for small object detection by complex dynamic attention mechanism of claim 1, wherein the downsampling and residual connection process of the encoder is: In the main branch, the real part characteristic and the imaginary part characteristic output by the dynamic fusion gating module are used for As input, respectively downsampling by convolution with preset step length to obtain real part characteristic and imaginary part characteristic after downsampling ; In the residual branch, the real part characteristic and the imaginary part characteristic of the input for complex residual block Convolution pairs using preset step sizes Processing Ji Cancha branches to obtain real part characteristics and imaginary part characteristics after residual branch alignment ; Downsampled real part characteristics Real part feature aligned with residual branch Adding and activating to obtain real part characteristic of encoder output Downsampled imaginary features Imaginary feature aligned with residual branches Adding and activating to obtain the imaginary part characteristics of the encoder output 。
8. The small object detection method of complex dynamic attention mechanism as in claim 1, wherein the decoder gradually restores the real and imaginary features extracted by the encoder to the original spatial dimensions by transposed convolution recovery resolution operation, in which process the real and imaginary features obtained by the encoder are combined with the upsampling result by a jump connection to fuse the multi-scale features to preserve more detailed information; The decoder performs up-sampling step by step and fuses with the jump connection feature of the encoder to obtain the decoded real part feature and the imaginary part feature which are in the same scale as the input ; After channel dimension splicing, mapping to three category channels through convolution, so as to generate three-category non-normalized scores; performing Softmax logistic regression on the un-normalized score to obtain a pixel Belonging to the first Prediction probability of class 。
9. The method for small object detection by complex dynamic attention mechanisms of claim 1, wherein the mixed Loss function combines Focal Loss and Tversky Loss; Focal Loss is defined as: Wherein the method comprises the steps of Representing the true class of the training sample; an effective mask for training samples; Is a pixel In the true category Predictive probability on; Is a modulation factor; Tversky Loss balancing the classification capability of the network to different classes by controlling the punishment proportion of false positives and false negatives, defining a binary truth-value diagram to the target class And predictive probability map Is combined with And (3) with Defining true positive cumulative values within a valid mask Cumulative value of false positives And false negative cumulative value Tversky Loss is defined as: Wherein, the Tversky Loss; To trade off the coefficients of false positives and false negatives; Is a normal number constant; The mixing loss function is: Wherein, the Is a weight coefficient.
10. Terminal device comprising a processor, a memory and a computer program stored in said memory, characterized in that the processor implements the small object detection method of the complex dynamic attention mechanism of any of claims 1-9 when executing the computer program.

Description

Small target detection method of complex dynamic attention mechanism Technical Field The invention relates to the field of radar signal processing and target detection, in particular to a small target detection method of a complex dynamic attention mechanism, which is based on deep learning, image segmentation and signal processing and realizes accurate detection and segmentation of small targets and clutter in a distance-Doppler domain. Background As an active microwave target detection device, the radar has all-day, all-weather and high-precision detection capability, and is widely applied to a plurality of fields such as military defense, autopilot, environmental perception and the like. The traditional radar target detection method mainly adopts the difference of target and clutter energy, on one hand, clutter is suppressed through moving target indication (Moving Target Indicator, MTI), on the other hand, moving target detection (Moving Target Detection, MTD) is adopted, multi-pulse combined accumulation is combined to realize target energy accumulation, and finally, a Constant false alarm detection method (Constant FALSE ALARM RATE, CFAR) is adopted, and statistic information is utilized to realize reliable detection of the target. However, due to the influence of the strong land and sea clutter, when the traditional method detects slow small targets such as unmanned aerial vehicles, the conditions of residual clutter residues, false alarms and missed alarms of the targets exist, and the detection performance of the radar is severely restricted. In recent years, the rapid development of deep learning technology provides a new solution for weak and small target detection. Such methods transform the traditional energy detection problem into an intelligent segmentation problem of a two-dimensional image domain, and common segmentation methods include convolutional neural Networks (Convolutional Neural Networks, CNN), generation of countermeasure Networks (GENERATIVE ADVERSARIAL Networks, GAN), and the like. The pixel level segmentation method based on the U-Net structure achieves remarkable results in medical image segmentation and natural image segmentation. Most of the existing deep learning methods focus on feature extraction by using amplitude information of echo signals, but ignore rich phase information contained in radar signals, i.e., I (in-phase) and Q (quadrature) components of complex domain. Phase information is of great importance in radar echo for describing weak features and details of targets, and is particularly critical when dealing with small targets and low-contrast targets. In addition, for the scheme adopting the super-resolution or fine-granularity segmentation network, the data labeling and model training cost is high, and in a practical complex environment, modeling is performed only by relying on single characteristic dimensions such as amplitude, phase and the like, so that the requirements of target detection and segmentation are difficult to be met at the same time. Disclosure of Invention The invention aims to provide a small target detection method of a complex dynamic attention mechanism, which aims to solve the problem of difficult detection of a small target under a complex clutter background. In order to realize the tasks, the invention adopts the following technical scheme: A method for small object detection for a complex dynamic attention mechanism, comprising: The method comprises the steps of taking a complex matrix obtained by pulse pressure and moving target detection processing of target reflection echoes received by a radar as input, establishing a background, clutter and target three-classification label graph as labels and setting an effective mask by constructing an I/Q (input/output) dual-channel standardized tensor as a training sample, constructing a small target detection and segmentation network based on U-Net in a training stage, embedding a complex domain dynamic attention module after a complex residual block of an encoder, wherein the complex domain dynamic attention module comprises a complex domain channel attention module, a multi-scale space attention module and a dynamic fusion gating module which are respectively used for carrying out channel enhancement and space enhancement on complex features and then carrying out dynamic fusion, adopting a mixing loss function in training, outputting a prediction category in an inference stage, and obtaining a detection result by a target category mask. Further, the small target detection and segmentation network takes U-Net as a main body and comprises an encoder, a decoder and an output head, wherein the encoder is formed by stacking a plurality of complex residual blocks, a complex domain dynamic attention module is embedded in the complex residual blocks, the equivalent complex representation of I/Q double channels is adopted, and the input complex characteristics and the output complex characteristics