PolSAR target detection method based on polarization characteristics and anomaly detection
By constructing an unsupervised PolSAR target detection network and utilizing polarization characteristics and anomaly detection, the problem of existing methods' dependence on labeled data is solved, and the performance of PolSAR target detection is improved, especially in ground vehicle target detection.
Patent Information
- Application Number
- CN202311778636.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-12-21
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2043-12-21
AI Technical Summary
Existing PolSAR target detection methods require a large amount of labeled data for training and do not fully utilize polarization scattering characteristics, making it difficult to achieve efficient adaptive detection in ground target detection.
A PolSAR target detection method based on polarization characteristics and anomaly detection is adopted. By acquiring the polarization correlation features and polarization coherence matrix of PolSAR images, an unsupervised PolSAR target detection network is constructed. Feature extraction and reconstruction are performed using an encoder, decoder, memory library module and discriminator. Combined with polarization rotation domain features, unsupervised target detection is carried out.
In the absence of a large amount of labeled data, the performance of PolSAR target detection has been improved, achieving efficient target detection, especially for ground vehicle targets.
Smart Images

Figure CN117710657B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of radar technology, specifically relating to a PolSAR target detection method based on polarization characteristics and anomaly detection. Background Technology
[0002] Space remote sensing systems include optical remote sensing systems (visible light remote sensing, thermal infrared remote sensing, etc.) and microwave remote sensing systems such as Synthetic Aperture Radar (SAR) systems and microwave radiometers. Traditional optical remote sensing systems can only operate under illumination and are easily affected by weather conditions such as clouds, rain, and fog. In contrast, SAR systems are active microwave imaging systems capable of all-weather Earth observation. Operating in the microwave band, SAR beams can effectively penetrate clouds and dense fog, enabling all-weather Earth observation. The electromagnetic waves used by SAR also have some penetrating power through vegetation and camouflage, allowing the acquisition of echoes from targets hidden under vegetation or in shallow subsurface areas. Therefore, SAR has unique advantages in military applications. Compared to single-polarization SAR systems, fully polarization SAR systems can acquire not only the radar cross-section (RCS) of a target but also the complete electromagnetic scattering characteristics (i.e., the polarization scattering matrix) of the observed target, enriching the dimensions of the acquired information. By utilizing polarimetric scattering matrices, researchers can analyze and extract the physical parameters and structural information of targets in greater depth. Therefore, the study of PolSAR (Polarimetric Synthetic Aperture Radar) images has significant theoretical and practical value.
[0003] In the late 1980s, researchers at Lincoln Laboratory began studying detection algorithms based on PolSAR data. Due to Lincoln Laboratory's advantage in developing advanced PolSAR systems, it acquired a large amount of valuable polarimetric SAR data. In 1989, Novak et al. proposed the Optimal Polarimetric Detector (OPD). Subsequently, Novak et al. proposed the Polarimetric Matched Filter Detector (PMF). Later, Novak and Burl proposed the Polarimetric Whitening Filter Detector (PWF). Marino et al. proposed the Polarimetric Notch Filter (PNF) with adaptive detection capabilities, successfully achieving the detection of ship targets on the sea surface. Nunziata et al. constructed a Reflection Symmetry (RS) detector for detecting ship targets in PolSAR images. Chen et al. proposed the polarimetric rotation domain theory and a polarimetric correlation feature for PolSAR data, successfully applying it to the detection of ship targets in PolSAR. Patent CN202310912719.1 (“A CFAR-guided PolSAR target detection method based on polarization characteristics and superpixel segmentation”) provides a CFAR-guided PolSAR target detection method based on polarization characteristics and superpixel segmentation. By calculating polarization rotation domain features and utilizing superpixel segmentation and CFAR detection results, it enhances the difference between target and background features, thereby improving the performance of PolSAR vehicle target detection. However, traditional target detection methods typically require manual feature calculation and selection, making adaptive detection difficult, and their target detection accuracy is relatively low. Therefore, traditional target detection algorithms are gradually failing to meet the demands.
[0004] In recent years, deep learning-based target detection methods have become a focus of research for many scholars, and have achieved considerable success in SAR ship target detection. However, research on using convolutional neural networks for PolSAR target detection is relatively limited, especially for ground target detection tasks. Cozzolino et al. used a small convolutional neural network to detect ships in SAR images, demonstrating that features extracted by the neural network are more suitable for detection than manually designed features. Nieto Hidalgo et al. proposed a two-stage target detection framework for specific backgrounds, which is highly efficient for marine target detection, but not for other scenarios (such as ground vehicle target detection). Fan Weiwei et al. used an improved Faster R-CNN network for PolSAR ship detection. Jin Kan et al. proposed a Patch to Pixel CNN (P2P-CNN) convolutional neural network for ship detection in PolSAR images, which significantly improved detection performance.
[0005] However, the above methods all require target-labeled data to train the network, and they do not make full use of polarization scattering characteristics. Summary of the Invention
[0006] To address the aforementioned problems in the existing technology, this invention provides a PolSAR target detection method based on polarization characteristics and anomaly detection.
[0007] The technical problem to be solved by this invention is achieved through the following technical solution:
[0008] A PolSAR target detection method based on polarization characteristics and anomaly detection, the PolSAR target detection method comprising:
[0009] Acquire the PolSAR image to be tested;
[0010] Test data is obtained by splicing the polarization correlation features and polarization coherence matrix of the PolSAR image to be tested.
[0011] The test data is sliced to obtain final test data including multiple test slices, with some overlap between adjacent test slices;
[0012] The test slice is input into a trained PolSAR target detection network to obtain reconstructed data for each test slice. The anomaly detection score map of the test slice is obtained by averaging the squared difference between the reconstructed data and the corresponding pixel of the test slice along the channel dimension. The PolSAR target detection network includes an encoder, a decoder, a memory library module, and a discriminator. The encoder is used to extract multiple features of the PolSAR image and inputs the extracted features into the memory library module. The memory library module is used for attention-based addressing and weighted summation to obtain multiple features. The multiple features output by the memory library module are input into the decoder. The decoder is used to reconstruct the input features to obtain reconstructed data. The discriminator is used to extract features from the input data and the reconstructed data.
[0013] The anomaly detection score map is mapped back to the PolSAR image to be tested, and segmented according to a threshold to obtain the final binary detection result.
[0014] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0015] This invention addresses the current lack of large amounts of labeled PolSAR data by proposing a method for PolSAR target detection using unsupervised anomaly detection. First, test data is obtained by splicing the polarization correlation features and polarization coherence matrix of the target region in the PolSAR image to be tested. Test slices of this data are then input into a trained PolSAR target detection network to obtain anomaly detection scores. Finally, the anomaly detection score map is mapped back to the PolSAR image to be tested and segmented according to a threshold to obtain the final binary detection result. The PolSAR target detection network of this invention consists of an encoder, decoder, memory library module, and discriminator, eliminating the need for extensive training with labeled PolSAR data. Furthermore, the PolSAR target detection network incorporates polarization rotation domain correlation features as input, further improving PolSAR target detection performance.
[0016] The present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description
[0017] Figure 1 This is a flowchart illustrating a PolSAR target detection method based on polarization characteristics and anomaly detection provided in an embodiment of the present invention.
[0018] Figure 2 This is a schematic diagram illustrating the implementation process of a PolSAR target detection method based on polarization characteristics and anomaly detection provided in an embodiment of the present invention.
[0019] Figure 3 This is a network framework diagram of a PolSAR target detection network provided in an embodiment of the present invention;
[0020] Figure 4 This is an anomaly detection score map of a test image provided in an embodiment of the present invention;
[0021] Figure 5 This is a final detection result diagram provided by an embodiment of the present invention;
[0022] Figure 6 This is a detection result image obtained by a method in a patented comparison algorithm provided in this invention embodiment;
[0023] Figure 7 This is a detection result image obtained by a comparison algorithm PWF provided in an embodiment of the present invention;
[0024] Figure 8 This is a detection result image obtained by a comparison algorithm RS provided in an embodiment of the present invention; Detailed Implementation
[0025] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.
[0026] Example 1
[0027] With the continuous deepening of SAR theoretical research and the ongoing advancement of system development, SAR systems are rapidly evolving towards multi-band, multi-view, and multi-polarization directions. In particular, Polarimetric Synthetic Aperture Radar (PolSAR), with its full polarization measurement capability, has attracted greater attention from various countries due to its superior characteristics and has become one of the mainstream research directions for scholars today. In recent years, with the great success of deep networks in the field of optical image target detection, deep learning-based PolSAR target detection methods have become a hot research topic for many scholars. While some progress has been made in PolSAR ship target detection tasks, ground target detection is extremely difficult due to the limited amount of PolSAR data and the scarcity of targets in the images, coupled with the presence of clutter with similar characteristics to the targets. Therefore, research on ground vehicle target detection is relatively limited. Existing deep network-based PolSAR target detection methods do not deeply consider the characteristics of PolSAR images themselves and the benefits of polarization scattering for target detection. Furthermore, most existing methods are supervised target detection network frameworks, requiring a large amount of labeled data for training to achieve good results. However, in reality, a large amount of labeled PolSAR data is generally unavailable. Therefore, existing algorithms still have certain limitations.
[0028] Based on this, the present invention provides a PolSAR target detection method based on polarization characteristics and anomaly detection. Please refer to [link to relevant documentation]. Figure 1 , Figure 1 This is a flowchart illustrating a PolSAR target detection method based on polarization characteristics and anomaly detection provided in an embodiment of the present invention. The PolSAR target detection method includes:
[0029] Step 1: Obtain the PolSAR image to be tested.
[0030] Step 2: Obtain test data based on the splicing results of the polarization correlation features and polarization coherence matrix of the PolSAR image to be tested.
[0031] Specifically, the PolSAR image to be tested is a region containing the target, for example, a size of 768*768. The polarization correlation features and polarization coherence matrix of the test image are calculated. From all the obtained polarization correlation features, N polarization correlation features that are the same as those selected during the training of the PolSAR target detection network (in the training phase, the top N polarization correlation features with higher signal-to-clutter ratios are selected). Then, the selected N polarization correlation features are concatenated with the elements of the polarization coherence matrix of the test image to obtain the test data.
[0032] Here, the method for concatenating the elements of the polarization coherence matrix of the test image with the first N polarization-related features is as follows: The upper triangular elements of the polarization coherence matrix of the test image have 6 elements, three of which are complex numbers. These are split into real and imaginary parts. In the channel dimension, the real numbers in the upper triangular elements of the polarization coherence matrix, as well as the split real and imaginary parts, are concatenated into a 9-dimensional matrix. Then, in the channel dimension, the 9-dimensional matrix is concatenated with the first N polarization-related features. The result of the concatenation is the test data. For example, if N is 6, a 15-dimensional matrix is obtained. Therefore, the original test data size is 768*768*15.
[0033] In this embodiment, a method for obtaining the polarization coherence matrix and polarization correlation features is also provided, specifically including:
[0034] A1. Construct the polarization scattering matrix S based on the four channels of the PolSAR image.
[0035]
[0036] Where S is the polarization scattering matrix, S HH To represent the SAR images corresponding to the transmission and reception of electromagnetic waves with horizontal polarization, S HV For SAR images corresponding to transmitting electromagnetic waves with horizontal polarization and receiving electromagnetic waves with vertical polarization, S VHFor SAR images corresponding to vertically polarized electromagnetic wave transmission and horizontally polarized electromagnetic wave reception, S VV These are SAR images corresponding to the transmission and reception of electromagnetic waves using vertical polarization.
[0037] A2. Based on the polarization scattering matrix S from step A1, calculate the polarization coherence matrix using the following formula:
[0038]
[0039] Where T is the polarization coherence matrix, ·* is the conjugate, and <·> is the sample geometric mean.
[0040] A3. As shown in step A1, for a PolSAR image, its polarization scattering matrix S can be obtained. Using a rotation matrix, the polarization scattering matrix is rotated along the radar line of sight at a preset angle to obtain the rotated polarization scattering matrix. The rotated polarization scattering matrix is expressed as:
[0041]
[0042]
[0043] Where S(θ) is the rotated polarization scattering matrix, θ is the preset angle, θ∈[-π,π), the superscript T is the transpose, and R2(θ) is the rotation matrix.
[0044] The four components of the rotated polarization scattering matrix are represented as follows:
[0045] S HH (θ)=S HH cos 2 θ+S HV cosθsinθ+S VH cosθsinθ+S VV sin 2 θ
[0046] S HV (θ)=-S HH cosθsinθ+S HV cos 2 θ-S VH sin 2 θ+S VV cosθsinθ
[0047] S VH (θ)=-S HH cosθsinθ-S HV sin 2 θ+S VH cos 2 θ+SVV cosθsinθ
[0048] S VV (θ)=S HH sin 2 θ-S HV cosθsinθ-S VH cosθsinθ+S VV cos 2 θ
[0049] Among them, S HH (θ) represents the rotated SAR images corresponding to the horizontally polarized transmission and reception of electromagnetic waves, S HV (θ) represents the rotated SAR images corresponding to the transmission and reception of electromagnetic waves with horizontal polarization and vertical polarization, respectively. VH (θ) represents the rotated SAR image corresponding to the vertically polarized transmission and horizontally polarized reception of electromagnetic waves. VV (θ) represents the rotated SAR image corresponding to the vertically polarized transmission and reception of electromagnetic waves.
[0050] A4. Under the condition that the reciprocity theorem S is satisfied... HV =S VH Under the condition of Pauli feature vectors, and Lexicographic feature vector Six polarization-related modes were obtained, including the first polarization-related mode |γ HH-VV (θ)|、Second polarization-related mode|γ HH-HV (θ)|、Third polarization related mode|γ VV-HV (θ)|、Fourth polarization related mode|γ (HH+VV)-(HH-VV) (θ)|、Fifth polarization-related mode|γ (HH+VV)-HV |and the sixth polarization-related mode|γ (HH-VV)-(HV) (θ)|.
[0051] Specifically, for the two polarization channels S1 and S2, the polarization correlation is defined as follows:
[0052]
[0053] in, The conjugate of S2 is represented by , and <·> represents the geometric mean of the samples. In PolSAR images, the correlation between the two polarization channels contains rich information. Extending this original polarization correlation between the two channels at a fixed angle to a rotational domain along the radar line of sight yields the following polarization correlation pattern:
[0054]
[0055] in, This represents the conjugate of S2(θ).
[0056] Therefore, under the premise of satisfying the reciprocity theorem, i.e., S HV =S VH Based on Pauli feature vectors and Lexicographic feature vector Six typical polarization-related modes can be obtained, namely |γ HH-VV (θ)|、|γ HH-HV (θ)|、|γ VV-HV (θ)|、|γ (HH+VV)-(HH-VV) (θ)|、|γ (HH+VV)-HV (θ)| and |γ (HH-VV)-HV (θ)|.
[0057] A5, based on |γ (HH+VV)-(HH-VV) (θ)|=|γ (HH+VV)-HV (θ+π / 4)| and |γ HH-VV (θ)|=|γ VV-HV (θ+π / 2)| yields four final polarization correlation modes, including |γ|. HH-VV (θ)|、γ HH-HV (θ)|、|γ (HH+VV)-(HH-VV) (θ)| and |γ (HH-VV)-(HV) (θ)|.
[0058] Here, because:
[0059]
[0060]
[0061] The following conclusions can be drawn:
[0062] |γ (HH+VV)-(HH-VV) (θ)|=|γ (HH+VV)-HV (θ+π / 4)|
[0063] |γ HH-VV (θ)|=|γ VV-HV (θ+π / 2)|
[0064] Therefore, this means |γ (HH+VV)-(HH-VV) (θ)| and |γ (HH+VV)-HV (θ)|、|γ HH-VV (θ)| and |γ VV-HV These two pairs of polarization correlations (θ) are equivalent. Therefore, four independent polarization correlation modes |γ are ultimately obtained. HH-VV (θ)|、|γ HH-HV(θ)|、|γ (HH+VV)-(HH-VV) (θ)| and |γ (HH-VV)-(HV) The specific calculation formula for (θ)| is shown below:
[0065]
[0066]
[0067]
[0068]
[0069] Among them, S (HH+VV) (θ) is S HH (θ) and S VV The sum of (θ), S (HH-VV) (θ) is S HH (θ) and S VV The difference (θ).
[0070] Therefore, four polarization correlation modes are calculated according to the above formula, where θ takes values from -π to π at intervals of 0.01, and the corresponding polarization correlation modes at different angles are calculated iteratively.
[0071] A6. For each polarization-related mode, calculate the signal-to-noise ratio (SNR) of each polarization-related feature in the target region and the background region.
[0072] In this embodiment, to more accurately quantitatively characterize polarization-related modes, a set of polarization-related features is used. These features are then calculated to further uncover hidden information in the rotation domain. The definitions of these polarization-related features are as follows:
[0073] 1) Original correlation feature γ -org This feature is a polarization correlation feature typically used without any rotation processing, representing the target decorrelation effect between the two polarization channels under the original imaging geometry.
[0074] 2) Correlation mean γ -mean It is the average correlation value in the rotation domain, representing the general level of correlation. The correlation average is a measure of the average decorrelation effect of the target in the rotation domain. The higher the correlation average, the lower the decorrelation phenomenon.
[0075] 3) Correlation standard deviation γ -std It is the correlation standard deviation in the rotation domain. It can identify the diversity of target scattering directions. Generally, the larger the correlation standard deviation, the more pronounced the diversity of target scattering directions in the rotation domain. For rotationally invariant scatterers, the coherence fluctuation will approach 0.
[0076] 4) Maximum correlation γ -max: Maximum value of rotational domain correlation, the upper limit of polarization correlation between two polarization channels at different angles.
[0077] 5) Minimal correlation γ -min : Minimum of rotational domain correlation, the lower limit of polarization correlation between two polarization channels at different angles.
[0078] 6) Correlation contrast γ -contrast =γ -max -γ -min It reflects the absolute contrast in the rotation domain. Furthermore, the correlation contrast can also reflect the diversity of target scattering directions in the rotation domain. For rotationally invariant targets with no direction dependence, the correlation contrast will become 0.
[0079] 7) Correlation anisotropy γ -A =(γ -max -γ -min ) / (γ -max +γ -min ): Reflects the relative contrast in the rotation domain. Correlation anisotropy is a complementary feature to correlation contrast. For rotating targets with relatively low correlation, correlation anisotropy can further enhance contrast and provide better discrimination performance.
[0080] 8) Correlated beamwidth γ -bw0.95 That is, the correlation value is not less than 0.95 × γ -max The smaller the value of the rotation angle range, the greater the decorrelation effect and the stronger the direction dependence.
[0081] 9) Maximum rotation angle θ γ-max Defined as generating the maximum relevant rotation angle within the main range.
[0082] 10) Minimum rotation angle θ γ-min Defined as generating the minimum relevant rotation angle within the main range.
[0083] Based on the definition of polarization-related features, the ten polarization-related features corresponding to each polarization-related mode are calculated, resulting in a total of 40 polarization-related features. These polarization-related features have a certain degree of redundancy.
[0084] For each of the four polarization-related modes, ten polarization-related features can be obtained as described above. Therefore, it is necessary to select features with significant differences between the target and background, and the selection will vary depending on the dataset. Thus, to select features with significant differences between the target and background, this embodiment calculates the signal-to-noise ratio (SNR) of each polarization-related feature corresponding to the target and background regions, and selects the N polarization-related features with the highest SNR from all polarization-related features corresponding to the target and background regions. For example, N is 6.
[0085] Step 3: Slice the test data to obtain the final test data, which includes multiple test slices, with some overlap between adjacent test slices.
[0086] For example, by using a step size of 64 and a slice size of 128*128, a sliding window slice of a 768*768 image is obtained to produce 36 slice images, thus yielding the final test data.
[0087] Step 4: Input the test slices into the trained PolSAR target detection network to obtain the reconstructed data of each test slice. Based on the square of the difference between the reconstructed data and the corresponding pixel points of the test slice, and average it in the channel dimension, the anomaly detection score map of the test slice is obtained.
[0088] In one specific embodiment, please refer to Figure 2 This embodiment provides a training method for a PolSAR target detection network, the training method including:
[0089] S1. Obtain the original PolSAR training images and target training samples.
[0090] S2. Select a region in the training image as the background region, calculate the signal-to-clutter ratio (SCR) of each polarization-related feature in the background region and the target training sample, and select the top N polarization-related features with the highest SCR for subsequent experiments. Calculate the above N polarization-related features in the original PolSAR training image and concatenate them with the elements of their polarization coherence matrix to form the original training data.
[0091] Specifically, the original PolSAR training image is a pure background region, for example, with a size of 1280*896, while the target training sample is a target region data obtained from manual annotation. A region is selected in the training image as the background region, and the signal-to-clutter ratio of each polarization-related feature in the background region and the target region is calculated respectively. The top N polarization-related features with the highest signal-to-clutter ratio are selected. Then, the above N polarization-related features of the original PolSAR training image are calculated and concatenated with the elements of their polarization coherence matrix to obtain the original training data.
[0092] Here, the method for concatenating the elements of the polarization coherence matrix with the first N polarization-related features is as follows: The upper triangular elements of the polarization coherence matrix of the background region have 6 elements, three of which are complex numbers. These are split into real and imaginary parts. In the channel dimension, the real numbers in the upper triangular elements of the polarization coherence matrix, as well as the split real and imaginary parts, are concatenated into a 9-dimensional matrix. Then, in the channel dimension, the 9-dimensional matrix is concatenated with the first N polarization-related features. The result of the concatenation is the original training data. For example, if N is 6, a 15-dimensional matrix is obtained. Therefore, the size of the original training data is 1280*896*15.
[0093] S3. The original training data is sliced to obtain the final training data, which includes multiple training slices, with some overlap between adjacent training slices.
[0094] For example, with a step size of 64 and a slice size of 128*128, a sliding window slice of a 1280*896 image is obtained to get 247 training slices, and these 247 training slices constitute the final training data ψ.
[0095] S4. Input the final training data into Module 1 of the PolSAR target detection network to obtain reconstructed data. During the training process, unsupervised learning is performed on the similarity between the final training data and the reconstructed data.
[0096] Please see Figure 3 This embodiment provides a specific structural diagram of a PolSAR target detection network. Module one includes an encoder, a decoder, and a memory library module. Specifically, as shown... Figure 3 As shown, the encoder is a feature extraction network, including encoder1, encoder2, encoder3, encoder4, and encoder5 modules. The encoder outputs are the output feature layers C2, C3, C4, and C5 of each module. The decoder is a feature reconstruction network, including decoder1, decoder2, decoder3, decoder4, and decoder5 modules. The input to each module of the decoder network is the concatenation of the output C′2, C′3, C′4, and C′5 of the corresponding encoder's output feature layer after passing through the memory library module, and the result of the previous layer. In this embodiment, attention-based addressing query-based block memory libraries are constructed after the output feature layers C2, C3, C4, and C5. The modules corresponding to each block memory library are mem1, mem2, mem3, and mem4, and each module contains a corresponding memory library M. j (j = 1, 2, 3, 4). The output feature layer C of the encoder... i (i = 2, 3, 4, 5) are used as the input to the module corresponding to the i-th layer block memory library. After attention-based addressing and querying, a weighted sum is obtained to obtain the new feature C′. i (i = 2, 3, 4, 5), C′ i This serves as the output of the module corresponding to the i-th level block memory library.
[0097] In this embodiment, the encoder includes encoder1, encoder2, encoder3, encoder4, and encoder5 modules. The input data size of the encoder is H×W×K, where H is the height, W is the width, and K is the channel dimension, for example, 128*128*15.
[0098] The encoder1 module includes a convolutional layer L1 and a convolutional layer L2. The convolutional layer L1 is used to output a1 feature maps of size H×W. The L1 convolutional layer outputs a1 feature maps The input is fed into convolutional layer L2, which outputs a1 feature maps of size H×W. L2 outputs a1 feature maps Input to the encoder2 module.
[0099] Specifically, the convolutional kernel in convolutional layer L1 has a window size of 3×3, a stride of 1, and a padding attribute of 1, and is used to output 64 feature maps of size H×W. Let L1 represent the j1-th feature map. The output of convolutional layer L1 serves as the input to convolutional layer L2. The convolutional kernel in convolutional layer L2 has a 3×3 window size, a stride of 1, and a padding attribute of 1, outputting 64 feature maps of size H×W. Let J represent the j1-th feature map. The output of the convolutional layer L2 is the input of the encoder2 module.
[0100] The encoder2 module includes a pooling layer P1, a convolutional layer L3, and a convolutional layer L4. The pooling layer P1 is used to output a1 units. Size feature map P1 outputs a1 feature maps The input is fed into convolutional layer L3, which outputs a2 values. Size feature map The L3 outputs a2 feature maps The input is fed into convolutional layer L4, which outputs a2 values. Size feature map L4 outputs a2 feature maps The input is fed into the encoder3 module, and the L4 outputs a2 feature maps. This is the encoder's output characteristic C2, where a2 = 2 * a1.
[0101] Specifically, pooling layer P1 uses max pooling, with a kernel window size of 2×2, a stride of 2, and a padding attribute of 1, for outputting 64... Size feature map Let $\mathbf{j}$ represent the j1-th feature map. The output of pooling layer P1 serves as the input to convolutional layer L3. The convolutional kernel in convolutional layer L3 has a 3×3 window size, a stride of 1, and a padding attribute of 1, and outputs 128 features. Size feature map Let L represent the j2-th feature map. The output of convolutional layer L3 serves as the input to convolutional layer L4. The convolutional kernel in convolutional layer L4 has a 3×3 window size, a stride of 1, and a padding attribute of 1, and outputs 128 features. Size feature map Let J represent the j2th feature map. The output of convolutional layer L4 is the input of encoder3 module. In addition, the output of L4 is the output feature C2 of the encoder.
[0102] The encoder3 module includes a pooling layer P2, a convolutional layer L5, and a convolutional layer L6. The pooling layer P2 is used to output a2 values. Size feature map P2 outputs a2 feature maps The input is fed into convolutional layer L5, which outputs a3 values. Size feature map The L5 outputs a3 feature maps The input is fed into convolutional layer L6, which outputs a3 values. Size feature map The L6 outputs a3 feature maps The input is fed into the encoder4 module, and the L6 convolutional layer outputs a3 feature maps. Let C3 be the output feature of the encoder, and a3 = 2 * a2.
[0103] Specifically, pooling layer P2 uses max pooling, with a kernel window size of 2×2, a stride of 2, and a padding attribute of 1, to output 128... Size feature map Let J represent the j2-th feature map. The output of pooling layer P2 serves as the input to convolutional layer L5. The convolutional kernel in convolutional layer L5 has a 3×3 window size, a stride of 1, and a padding attribute of 1, and outputs 256 features. Size feature map Let L represent the j3rd feature map. The output of convolutional layer L5 serves as the input to convolutional layer L6. The convolutional kernel in convolutional layer L6 has a 3×3 window size, a stride of 1, and a padding attribute of 1, and outputs 256 features. Size feature map Let J represent the j3rd feature map. The output of convolutional layer L6 is the input of encoder4 module. In addition, the output of convolutional layer L6 is the output feature C3 of the encoder.
[0104] The encoder4 module includes a pooling layer P3, a convolutional layer L7, and a convolutional layer L8. The pooling layer P3 is used to output a3 values. Size feature map The P3 outputs a3 feature maps The input is fed into convolutional layer L7, which outputs a4 values. Size feature map The L7 outputs a4 feature maps The input is fed into convolutional layer L8, which outputs a4 values. Size feature map The L8 outputs a4 feature maps The input is fed into the encoder5 module, and the L8 convolutional layer outputs a4 feature maps. Let C4 be the output feature of the encoder, and a4 = 2 * a3.
[0105] Specifically, pooling layer P3 uses max pooling, with a kernel window size of 2×2, a stride of 2, and a padding attribute of 1, to output 256... Size feature map This represents the j3rd feature map. The output of pooling layer P3 serves as the input to convolutional layer L7. The convolutional kernel in convolutional layer L7 has a 3×3 window size, a stride of 1, and a padding attribute of 1, and outputs 512 features. Size feature map This represents the j4th feature map. The output of convolutional layer L7 serves as the input to convolutional layer L8. The convolutional kernel in convolutional layer L8 has a 3×3 window size, a stride of 1, and a padding attribute of 1, and outputs 512 features. Size feature map Let J represent the j4th feature map. The output of convolutional layer L8 is the input of encoder5 module. In addition, the output of convolutional layer L8 is the output feature C4 of the encoder.
[0106] The encoder5 module includes a pooling layer P4, a convolutional layer L9, and a convolutional layer L1. 10 Pooling layer P4 is used to output a4 values. Size feature map The P4 outputs a4 feature maps The input is fed into convolutional layer L9, which outputs a4 values. Size feature map The L9 outputs a4 feature maps Input to convolutional layer L 10 Convolutional layer L 10 Used to output 4 Size feature map L 10 Output a4 feature maps C5 is the output feature of the encoder.
[0107] Specifically, pooling layer P4 uses max pooling, with a kernel window size of 2×2, a stride of 2, and a padding attribute of 1, to output 512... Size feature map This represents the j4th feature map. The output of pooling layer P4 serves as the input to convolutional layer L9. The convolutional kernel in convolutional layer L9 has a 3×3 window size, a stride of 1, and a padding attribute of 1, and outputs 512 features. Size feature map This represents the j4th feature map, and the output of convolutional layer L9 is the output of convolutional layer L. 10 The input. Convolutional layer L 10 The convolutional kernel has a 3×3 window size, a stride of 1, and a padding attribute of 1, and is used to output 512 [convolutional values]. Size feature map This represents the j4th feature map, and the convolutional layer L... 10 The output is the encoder's output feature C5.
[0108] In this embodiment, the memory library module includes a block-based mem1 module, a mem2 module, a mem3 module, and a mem4 module.
[0109] Here, the mem1 module takes the encoder's output feature C2 as input. First, it performs a block operation on C2 in the spatial dimension, dividing it evenly into a1 blocks of equal size. Each block is then flattened into a vector, resulting in a1 query vectors. For each query vector Calculate its relationship with the row vector in memory library M1. The similarity between them yields a set of attention weights. And reset the attention weights below the preset threshold to 0, and then use this set of attention weights. We weight the row vectors in memory library M1 and sum the weighted vectors to obtain a new feature vector. Then, based on the above a1 query vectors... We can obtain a1 new feature vectors. Finally, the above new feature vectors are... Reassembling yields a new set of feature maps of the same size as C2. This set of feature maps This is the output feature C′2 of the mem1 module.
[0110] Specifically, the input to the mem1 module is the encoder's output feature C2, with a size of First, after a block-based operation, the input output feature C2 is uniformly divided into 64 blocks in the spatial dimension, with each block having a size of [size missing]. Flatten each block into a vector, resulting in 64. Query vector of size This represents the j1-th query vector; thereafter, for each query vector, an addressing operation is performed in the memory library M1, which is of size M1. The matrix, the vector in the i-th row of memory library M1 is represented as... Addressing operations are performed by calculation and The similarity between them yields a set of attention weights. For the i-th attention weight, to ensure sparsity, it will be less than 1. The attention weights are reset to 0, thus adjusting the attention weights based on the query vector. For vectors By weighting the vectors and summing them, a new feature vector is obtained. Therefore, based on the 64 query vectors mentioned above, 64 new feature vectors can be obtained. eigenvectors of size This represents the j1-th eigenvector; finally, the 64 new eigenvectors are reassembled to obtain 128 new eigenvectors. Size feature map (That is, to restore the 64 new feature vectors to the same size as the input C2). This represents the j2th feature map. The new 128 feature maps are the output of the mem1 module, which is also the output feature C′2.
[0111] Here, the mem2 module takes the encoder's output feature C3 as input. First, it performs a block operation on C3 in the spatial dimension, dividing it evenly into a1 blocks of the same size. Each block is then flattened into a vector, resulting in a1 query vectors. For each query vector Calculate its relationship with the row vector in memory library M2. The similarity between them yields a set of attention weights. The attention weights below a preset threshold are reset to 0. Then, these attention weights are used to weight the row vectors in memory database M2, and the weighted vectors are summed to obtain a new feature vector. Based on the aforementioned a1 query vectors... We can obtain a1 new feature vectors. Finally, the above new feature vectors are... Reassembling yields a new set of feature maps of the same size as C3. This set of feature maps This is the output feature C′3 of the mem2 module.
[0112] Specifically, the input to the mem2 module is the encoder's output feature C3, with a size of First, after a block-based operation, the input output feature C3 is uniformly divided into 64 blocks in the spatial dimension, with each block having a size of [size missing]. Flatten each block into a vector, resulting in 64. Query vector of size This represents the j1-th query vector; thereafter, for each query vector, an addressing operation is performed in the memory database M2, which is of size M2. The matrix, the vector in the i-th row of memory library M2 is represented as Addressing operations are performed by calculation and The similarity between them yields a set of attention weights. For the i-th attention weight, to ensure sparsity, it will be less than 1. The attention weights are reset to 0, thus adjusting the attention weights based on the query vector. For vectors By weighting the vectors and summing them, a new feature vector is obtained. Therefore, based on the 64 query vectors mentioned above, 64 new feature vectors can be obtained. eigenvectors of size This represents the j1-th eigenvector; finally, the 64 new eigenvectors are reassembled to obtain 256 new eigenvectors. Size feature map (That is, to restore the 64 new feature vectors to the same size as the input C3). This represents the j3rd feature map. The new 256 feature maps are the output of the mem2 module, which is also the output feature C′3.
[0113] Here, the mem3 module takes the encoder's output feature C4 as input. First, it performs a block operation on C4 in the spatial dimension, dividing it evenly into a1 blocks of equal size. Each block is then flattened into a vector, resulting in a1 query vectors. For each query vector Calculate its relationship with the row vector in memory library M3. The similarity between them yields a set of attention weights. And reset the attention weights below the preset threshold to 0, and then use this set of attention weights. We weight the row vectors in memory library M3 and sum the weighted vectors to obtain a new feature vector. Then, based on the above a1 query vectors... We can obtain a1 new feature vectors. Finally, the new feature vectors are reassembled to obtain a new set of feature maps of the same size as C4. This set of feature maps This is the output feature C′4 of the mem3 module.
[0114] Specifically, the input to the mem3 module is the encoder's output feature C4, with a size of First, after a block-based operation, the input output feature C4 is uniformly divided into 64 blocks in the spatial dimension, with each block having a size of [size missing]. Flatten each block into a vector, resulting in 64. Query vector of size This represents the j1-th query vector; thereafter, for each query vector, an addressing operation is performed in memory library M3, which is of size M3. The matrix, the vector in the i-th row of memory library M3 is represented as Addressing operations are performed by calculation and The similarity between them yields a set of attention weights. For the i-th attention weight, to ensure sparsity, it will be less than 1. The attention weights are reset to 0, thus adjusting the attention weights based on the query vector. For vectors By weighting the vectors and summing them, a new feature vector is obtained. Therefore, based on the 64 query vectors mentioned above, 64 new feature vectors can be obtained. eigenvectors of size This represents the j1-th eigenvector; finally, the 64 new eigenvectors are reassembled to obtain 512 new eigenvectors. Size feature map (That is, to restore the 64 new feature vectors to the same size as the input C4). This represents the j4th feature map. The new 512 feature maps are the output of the mem3 module, which is also the output feature C′4.
[0115] Here, the mem4 module takes the encoder's output feature C5 as input. First, it performs a block operation on C5 in the spatial dimension, dividing it evenly into a1 blocks of equal size. Each block is then flattened into a vector, resulting in a1 query vectors. For each query vector, calculate its relationship with the row vector in memory library M4. The similarity between them yields a set of attention weights. And reset the attention weights below the preset threshold to 0, and then use this set of attention weights. We weight the row vectors in memory library M4 and sum the weighted vectors to obtain a new feature vector. Then, based on the above a1 query vectors... We can obtain a1 new feature vectors. Finally, the above new feature vectors are... Reassembling yields a new set of feature maps of the same size as C5. This set of feature maps This is the output feature C′5 of the mem4 module.
[0116] Specifically, the input to the mem4 module is the encoder's output feature C5, with a size of First, after a block-based operation, the input output feature C5 is uniformly divided into 64 blocks in the spatial dimension, with each block having a size of [size missing]. Flatten each block into a vector, resulting in 64. Query vector of size This represents the j-th query vector; subsequently, for each query vector, an addressing operation is performed in memory library M4, which is of size M4. The matrix, the vector in the i-th row of memory library M4 is represented as Addressing operations are performed by calculation and The similarity between them yields a set of attention weights. For the i-th attention weight, to ensure sparsity, it will be less than 1. The attention weights are reset to 0, thus adjusting the attention weights based on the query vector. For vectors By weighting the vectors and summing them, a new feature vector is obtained. Therefore, based on the 64 query vectors mentioned above, 64 new feature vectors can be obtained. eigenvectors of size This represents the j1-th eigenvector; finally, the 64 new eigenvectors are reassembled to obtain 512 new eigenvectors. Size feature map (That is, to restore the 64 new feature vectors to the same size as the input C5). This represents the j4th feature map. The new 512 feature maps are the output of the mem4 module, which is also the output feature C′5.
[0117] Optionally, the initialization method for memory libraries M1, M2, M3, and M4 is as follows: take the square root of the width of each memory library matrix and then take the reciprocal to obtain a number x. Then, uniformly sample a set of numbers from -x to x to initialize the memory library matrix, thereby obtaining the initial memory library matrix for each memory library matrix.
[0118] Optionally, for the attention weights of the l-th layer The calculation formula is as follows:
[0119]
[0120] in, For the j1st query vector in the l-th layer, Let be the vector in the i-th row of the memory library at level l. Let k be the vector in the k-th row of the memory library at level l, with superscript... T This indicates transpose.
[0121] Optionally, the attention weights calculated above can be used. right We then weight and sum the features to obtain the new feature vector of the l-th layer. The specific calculation formula is as follows:
[0122]
[0123] in, This is the j1-th feature vector of the l-th layer.
[0124] Optionally, the preset threshold is:
[0125] In this embodiment, the decoder includes decoder1 module, decoder2 module, decoder3 module, decoder4 module and decoder5 module, which correspond one-to-one with each layer of the encoder. From bottom to top, they are decoder5 module, decoder4 module, decoder3 module, decoder2 module and decoder1 module.
[0126] Here, the decoder5 module includes an upsampling layer U1 and a convolutional layer L. 11 and convolutional layer L 12 The input to the upsampling layer U1 is the output feature C′5. The upsampling layer U1 first performs bilinear interpolation on the output feature C′5 to obtain a4 features. Size feature map Then, the above four feature maps a4 are analyzed. Perform a convolution operation to obtain a4 values. Size feature map The a4 feature maps output by U1 Concatenating with the output feature C′4 yields a5 features. Size feature map The spliced a5 feature maps Input to convolutional layer L 11 Convolutional layer L 11 Used to output 4 Size feature map L 11 Output a4 feature maps Input to convolutional layer L 12 Convolutional layer L 12 Used to output 4 Size feature map a5 = 2 * a4.
[0127] Specifically, the input to the upsampling layer U1 is the output feature C′5, which is first subjected to bilinear interpolation to obtain 512 features. Size feature map This represents the j4th feature map; then, a convolution operation is performed with a kernel window size of 3×3, a stride of 1, and padding of 1, resulting in 512 features. Size feature map This represents the j4th feature map. The output of the upsampling layer U1 is concatenated with the output feature C′4 to obtain 1024 features. Size feature map As a convolutional layer L 11 Input, This represents the j5th feature map. Convolutional layer L 11 The convolutional kernel has a 3×3 window size, a stride of 1, and a padding attribute of 1, and is used to output 512 [convolutional values]. Size feature map This represents the j4th feature map, and the convolutional layer L... 11 The output of the convolutional layer L 12 The input. Convolutional layer L12 The convolutional kernel has a 3×3 window size, a stride of 1, and a padding attribute of 1, and is used to output 512 [convolutional values]. Size feature map This represents the j4th feature map, and the convolutional layer L... 12 The output is a portion of the input to the decoder4 module.
[0128] Here, the decoder4 module includes an upsampling layer U2 and a convolutional layer L. 13 and convolutional layer L 14 The input to the upsampling layer U2 is L 12 The output feature map is first subjected to bilinear interpolation by the upsampling layer U2 to obtain a4 features. Size feature map Then, the above four feature maps a4 are analyzed. Perform a convolution operation to obtain a3 values. Size feature map The a3 feature maps output by U2 Concatenating with the output feature C′3 yields a4 features. Size feature map The stitched a4 feature maps Input to convolutional layer L 13 Convolutional layer L 13 Used to output a3 Size feature map L 13 Output a3 feature maps Input to convolutional layer L 14 Convolutional layer L 14 Used to output a3 Size feature map
[0129] Specifically, the input to the upsampling layer U2 is L 12 The output of the input upsampling layer U2 feature map is first bilinearly interpolated to obtain 512 features. Size feature map This represents the j4th feature map; then, a convolution operation is performed with a kernel window size of 3×3, a stride of 1, and padding of 1, resulting in 256 features. Size feature map This represents the j3rd feature map. The output of the upsampling layer U2 is concatenated with the output feature C′3 to obtain 512 features. Size feature map As a convolutional layer L 13 Input, This represents the j-4th feature map. Convolutional layer L 13 The convolutional kernel has a 3×3 window size, a stride of 1, and a padding attribute of 1, and is used to output 256 words. Size feature map Represents the j3rd feature map, convolutional layer L 13 The output of the convolutional layer L 14 The input. Convolutional layer L 14 The convolutional kernel has a 3×3 window size, a stride of 1, and a padding attribute of 1, and is used to output 256 words. Size feature map Represents the j3rd feature map, convolutional layer L 14 The output is a portion of the input to the decoder3 module.
[0130] Here, the decoder3 module includes an upsampling layer U3 and a convolutional layer L. 15 and convolutional layer L 16 The input to the upsampling layer U3 is L 14 The output feature map is first subjected to bilinear interpolation by the upsampling layer U3 to obtain a3 features. Size feature map Then, the above three feature maps a3 are analyzed. Perform a convolution operation to obtain a2 values. Size feature map The a2 feature maps output by U3 Concatenating with the output feature C′2 yields a3 features. Size feature map The spliced a3 feature maps Input to convolutional layer L 15 Convolutional layer L 15 Used to output a2 Size feature map L 15 Output a2 feature maps Input to convolutional layer L 16 Convolutional layer L 16 Used to output a2 Size feature map
[0131] Specifically, the input to the upsampling layer U3 is L 14 The output of the input upsampling layer U3 feature map is first bilinearly interpolated to obtain 256 features. Size feature map This represents the j3rd feature map; then, a convolution operation is performed with a kernel window size of 3×3, a stride of 1, and padding of 1, resulting in 128 features. Size feature map This represents the j2-th feature map. The output of the upsampling layer U3 is concatenated with the output feature C′2 to obtain 256 features. Size feature map As a convolutional layer L 15 Input, This represents the j3rd feature map. Convolutional layer L 15 The convolutional kernel has a 3×3 window size, a stride of 1, and a padding attribute of 1, and is used to output 128... Size feature map Represents the j2-th feature map, convolutional layer L 15 The output of the convolutional layer L 16 The input. Convolutional layer L 16 The convolutional kernel has a 3×3 window size, a stride of 1, and a padding attribute of 1, and is used to output 128... Size feature map Represents the j2-th feature map, convolutional layer L 16 The output is the input of the decoder2 module.
[0132] Here, the decoder2 module includes an upsampling layer U4 and a convolutional layer L. 17 The input to the upsampling layer U4 is L 16 The output feature map is first subjected to bilinear interpolation by the upsampling layer U4 to obtain a2 H×W feature maps. Then, the above a2 feature maps Perform convolution operations to obtain a1 feature maps of size H×W. The a1 feature maps output by U4 Input to convolutional layer L 17 Convolutional layer L 17 Used to output a1 feature maps of size H×W
[0133] Specifically, the input to the upsampling layer U4 is L 16 The output, input to the feature map of the upsampling layer U4, is first subjected to bilinear interpolation to obtain a feature map of size 128 H×W. This represents the j2-th feature map; then, a convolution operation is performed with a kernel window size of 3×3, a stride of 1, and a padding attribute of 1, resulting in 64 feature maps of size H×W. This represents the j1-th feature map, and the output of the upsampling layer U4 is the output of the convolutional layer L. 17 The input. Convolutional layer L 17 The convolutional kernel has a 3×3 window size, a stride of 1, and a padding attribute of 1, and outputs 64 feature maps of size H×W. Let L represent the j1-th feature map, and the convolutional layer L... 17 The output is the input of the decoder1 module.
[0134] Here, the decoder1 module includes convolutional layers L 18 L 17 The output feature map is input into the convolutional layer L 18 Convolutional layer L 18 Used to output 6 H×W feature maps L 18 Output a6 feature maps As reconstructed data.
[0135] Specifically, convolutional layer L 18 The convolutional kernel has a 3×3 window size, a stride of 1, and a padding attribute of 1, and outputs 15 feature maps of size H×W. This represents the j6th feature map, and the convolutional layer L... 18 The output is the reconstructed data ψ′.
[0136] S5. Obtain the first high-dimensional feature based on the final training data, and obtain the second high-dimensional feature based on the reconstructed data.
[0137] In this embodiment, the PolSAR target detection network further includes a discriminator, wherein:
[0138] The final training data is input into the discriminator to obtain the first high-dimensional feature;
[0139] The reconstructed data is input into the discriminator to obtain the second high-dimensional feature.
[0140] The discriminator in this embodiment is used to align the final training data and reconstructed data at the feature level. The discriminator is mainly composed of convolutional modules. The inputs to the discriminator are the final training data ψ and the reconstructed data ψ′, respectively, and the outputs are the corresponding first high-dimensional feature F and second high-dimensional feature F′. During training, unsupervised learning is performed on the similarity between the first high-dimensional feature F and the second high-dimensional feature F′.
[0141] Please see Figure 3 The present invention also provides a network structure for a discriminator, wherein the input data to the discriminator are the final training data ψ and the reconstructed data ψ′, both of which are of size H×W×15. The discriminator includes conv1 module, conv2 module, conv3 module, conv4 module, conv5 module and conv6 module.
[0142] Here, the conv1 module includes convolutional layers L 19 The data to be judged is input into the convolutional layer L. 19 Convolutional layer L 19 Used to output a1 Size feature map The data to be judged includes the final training data or the reconstructed data.
[0143] Specifically, convolutional layer L 19 The convolutional kernel has a window size of 4×4, a stride of 2, and a padding attribute of 1, and is used to output 64... Size feature map Let L represent the j1-th feature map, and the convolutional layer L... 19 The output is the input of the conv2 module.
[0144] Here, the conv2 module includes convolutional layers L 20 L 19 The output feature map is input into the convolutional layer L 20 Convolutional layer L 20 Used to output a2 Size feature map
[0145] Specifically, convolutional layer L 20 The convolutional kernel has a 4×4 window size, a stride of 2, and a padding attribute of 1, and is used to output 128... Size feature map Represents the j2-th feature map, convolutional layer L 20 The output is the input of the conv3 module.
[0146] Here, the conv3 module includes convolutional layers L 21 L 20 The output feature map is input into the convolutional layer L 21 Convolutional layer L 21 Used to output a3 Size feature map
[0147] Specifically, convolutional layer L21 The convolutional kernel has a 4×4 window size, a stride of 2, and a padding attribute of 1, and is used to output 256 words. Size feature map Represents the j3rd feature map, convolutional layer L 21 The output is the input of the conv4 module.
[0148] Here, the conv4 module includes convolutional layers L 22 L 21 The output feature map is input into the convolutional layer L 22 Convolutional layer L 22 Used to output 4 Size feature map
[0149] Specifically, convolutional layer L 22 The convolutional kernel has a 4×4 window size, a stride of 2, and a padding attribute of 1, and is used to output 512 [convolutional values]. Size feature map This represents the j4th feature map, and the convolutional layer L... 22 The output is the input of the conv5 module.
[0150] Here, the conv5 module includes convolutional layers L 23 L 22 The output feature map is input into the convolutional layer L 23 Convolutional layer L 23 Used to output a5 Size feature map
[0151] Specifically, convolutional layer L 23 The convolutional kernel has a 4×4 window size, a stride of 2, and a padding attribute of 1, and is used to output 1024... Size feature map This represents the j5th feature map, and the convolutional layer L... 23 The output is the input of the conv6 module.
[0152] Here, the conv6 module includes convolutional layers L 24 L 23 The output feature map is input into the convolutional layer L 24 Convolutional layer L 24 Used to output a2 Size feature map L 24The output a2 feature maps are used as high-dimensional features (the final training data and reconstructed data are respectively input into the convolutional layer L of the discriminator). 19 After passing through the discriminator, the final output corresponds to the first high-dimensional feature F and the second high-dimensional feature F′.
[0153] Specifically, convolutional layer L 24 The convolutional kernel has a 4×4 window size, a stride of 1, and a padding attribute of 0, and is used to output 128... Size feature map L represents the j2-th feature map. 24 The output is the final high-dimensional feature.
[0154] S6. Based on the loss function, minimize the difference between the final training data and the reconstructed data, and minimize the difference between the first high-dimensional feature and the second high-dimensional feature to obtain the trained PolSAR target detection network.
[0155] During training, unsupervised learning is required to assess the similarity between the reconstructed data and the final training data, with loss functions including L1 Loss and SSIM Loss. Unsupervised learning is also required to assess the similarity between the first high-dimensional feature and the second high-dimensional feature, with loss function including MSE Loss. Therefore, the loss functions of the PolSAR target detection network include L1 Loss, SSIM Loss, and MSE Loss.
[0156] The L1Loss function calculates the absolute value of the difference between the final training data and the reconstructed data pixel by pixel, sums them, and then averages the results. Its expression is as follows:
[0157]
[0158] Where L_L1 represents the L1Loss function, P represents the set of pixel coordinates, p represents the point coordinates, x represents the final training data, y represents the reconstructed data, and N represents the total number of pixels.
[0159] The SSIM Loss function measures the similarity between two images based on three aspects: brightness, contrast, and structural information. This is more in line with human visual perception. Its expression is as follows:
[0160]
[0161] Where, μ x and Let μ represent the mean and variance of x, respectively. y and Let σ represent the mean and variance of y, respectively.xy C1 and C2 represent the covariance of x and y, and are constants to avoid instability when the SSIM value is 0.
[0162] In practical applications, a sliding window is typically used to calculate the mean and standard deviation of pixels within each pixel of an image, thus obtaining the SSIM value for that pixel. The expression for the SSIM Loss function is as follows:
[0163]
[0164] Here, L_SSIM represents the SSIM Loss function.
[0165] The MSE Loss function calculates the mean squared error by summing the squares of the differences between the first high-dimensional feature F and the second high-dimensional feature F′ point by point, and then taking the average. Its expression is as follows:
[0166]
[0167] Where L_MSE represents the MSE Loss function, P represents the set of pixel coordinates of the feature map in the high-dimensional feature, and p represents the pixel coordinates of the feature map in the high-dimensional feature.
[0168] Step 5: Map the anomaly detection score map back to the PolSAR image to be tested, and segment it according to the threshold to obtain the final binary detection result.
[0169] Step 5.1: Map the anomaly detection score map of each test slice back to the location of the PolSAR image to be tested. For the non-overlapping areas of two adjacent test slices, take the score of the corresponding anomaly detection score map. For the overlapping areas of two adjacent test slices, take the larger value of the scores in the anomaly detection score maps of the two adjacent test slices for the score of each pixel in the overlapping area. This will give you the final anomaly detection score map of the PolSAR image to be tested.
[0170] In other words, the anomaly detection score map of each test slice is mapped back to the location of the PolSAR image to be tested. Since there will be some overlap between two adjacent test slices, for the non-overlapping parts, the score of each pixel in the PolSAR image to be tested is the score of the anomaly detection score map of the corresponding test slice. For the overlapping parts of two adjacent test slices, the score of each pixel in the PolSAR image to be tested in the overlapping parts is the larger value of the scores in the anomaly detection score maps of the two test slices. In this way, the score of each pixel in the PolSAR image to be tested can be obtained. All scores constitute the final anomaly detection score map of the PolSAR image to be tested.
[0171] Step 5.2: Determine the relationship between the final score map of anomaly detection and the threshold th. If the score in the final score map of anomaly detection is greater than the threshold th, then set the value of the corresponding pixel in the PolSAR image to be tested to 1. If the score in the final score map of anomaly detection is less than or equal to the threshold th, then set the value of the corresponding pixel in the PolSAR image to be tested to 0, and obtain the final binary detection result.
[0172] Optionally, the threshold th can be set to 5.
[0173] This invention addresses the problem of limited PolSAR image data and lack of real-world annotations by applying the concept of anomaly detection to target detection tasks. Sparse targets in the image are considered as anomalous regions, while large background areas are considered as normal samples. An unsupervised reconstruction network is constructed, which is trained using only normal samples during the training phase and then fed into the trained network with samples containing targets during the testing phase. Since only normal samples are used during training, the network cannot reconstruct the anomalous regions, i.e., the target regions, effectively, thus completing the target detection task.
[0174] This invention addresses the problem of blurred image details in the original network by replacing the original adversarial loss function (the loss function used in the original generative adversarial network) and MSE loss function with L1 loss function and SSIM loss function. This ensures that the reconstruction network can further improve its ability to reconstruct image details, while also being more in line with human visual perception, thereby achieving better object detection performance.
[0175] This invention addresses the issue that targets exhibit different scattering characteristics depending on their attitude relative to the radar line of sight. It not only utilizes traditional polarimetric coherence matrix elements for the input data but also incorporates features from polarimetric correlation modes, providing the network with more useful information and making it more suitable for PolSAR target detection tasks, thereby further improving the performance of PolSAR target detection.
[0176] This embodiment selects a set of PolSAR image data containing vehicle targets, and the experimental data is imaged using the full polarization method.
[0177] Table 1 below shows the vehicle target detection results of the method of the present invention on the above dataset, and compares them with existing PolSAR image target detection algorithms: Polarimetric Whitening Filter Detector (PWF, from the paper "Optimal polarimetric processing for enhanced target detection[J].IEEE Transactions on Aerospace and Electronic Systems,1993,29(1):234-244"), Reflection Symmetry-based Target Detector RS (Reflection Symmetry, from the paper "Reflection symmetry for polarimetric observation of man-made metallic targets at sea[J].IEEE Journal of Oceanic Engineering,2012,37(3):384-394"), and patent CN202310912719.1 ("A CFAR-guided PolSAR target detection method based on polarization characteristics and superpixel segmentation"). The target-level detection rate P in Table 1 is shown in the table. d The false alarm rate P represents the percentage of correctly detected vehicle targets out of the total number of real targets, expressed at the pixel level. f This indicates the proportion of detected erroneous pixels to the total number of background pixels; Figure 4 This is the anomaly detection score map for the test image. Figure 5 The final test result image is shown below. Figure 6 , Figure 7 , Figure 8 The images show the detection results of the proposed algorithm, PWF, and RS, respectively. The results demonstrate that the proposed method exhibits fewer false alarms and better detection performance.
[0178] Table 1 Comparison of target detection performance between the method of the present invention and existing methods
[0179]
[0180] This invention addresses the shortcomings of existing PolSAR target detection algorithms by proposing a PolSAR target detection method based on polarization characteristics and anomaly detection. First, to address the limited availability of PolSAR data, an unsupervised reconstruction network framework is used, leveraging anomaly detection principles, to accomplish the target detection task. Second, in addition to using the PolSAR data itself, relevant features of the polarimetric coherence matrix are incorporated as input to the network, fully utilizing the inherent characteristics of PolSAR images. This method, through the combination of polarimetric scattering characteristics and deep neural networks, significantly improves the performance of PolSAR image vehicle target detection algorithms.
[0181] It should be noted that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined as "first" or "second" may explicitly or implicitly include one or more features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.
[0182] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features or characteristics described may be combined in any suitable manner in one or more embodiments or examples. In addition, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.
[0183] Although the invention has been described herein in conjunction with various embodiments, those skilled in the art will understand and implement other variations of the disclosed embodiments by reviewing the accompanying drawings and the disclosure in carrying out the claimed invention. In this specification, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality. While certain measures are described in different embodiments, this does not mean that these measures cannot be combined to produce good results.
[0184] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.
Claims
1. A PolSAR target detection method based on polarization characteristics and anomaly detection, characterized in that, The PolSAR target detection method includes: Acquire the PolSAR image to be tested; Test data is obtained by splicing the polarization correlation features and polarization coherence matrix of the PolSAR image to be tested. The test data is sliced to obtain final test data including multiple test slices, with some overlap between adjacent test slices; The test slice is input into a trained PolSAR target detection network to obtain reconstructed data for each test slice. The anomaly detection score map of the test slice is obtained by averaging the squared difference between the reconstructed data and the corresponding pixel of the test slice along the channel dimension. The PolSAR target detection network includes an encoder, a decoder, a memory library module, and a discriminator. The encoder is used to extract multiple features of the PolSAR image and inputs the extracted features into the memory library module. The memory library module is used for attention-based addressing and weighted summation to obtain multiple features. The multiple features output by the memory library module are input into the decoder. The decoder is used to reconstruct the input features to obtain reconstructed data. The discriminator is used to extract features from the input data and the reconstructed data. The anomaly detection score map is mapped back to the PolSAR image to be tested, and segmented according to a threshold to obtain the final binary detection result.
2. The PolSAR target detection method based on polarization characteristics and anomaly detection according to claim 1, characterized in that, The training method for the PolSAR target detection network includes: Acquire the original PolSAR training images and target training samples; In the original PolSAR training image, a region is selected as the background region. The signal-to-clutter ratio (SCR) of each polarization-related feature of the background region and the target training sample is calculated. The top N polarization-related features with higher SCRs are selected for subsequent experiments. The above N polarization-related features of the original PolSAR training image are calculated and concatenated with the elements of their polarization coherence matrix to form the original training data. The original training data is sliced to obtain the final training data, which includes multiple training slices, with some overlap between adjacent training slices. The final training data is input into module one of the PolSAR target detection network to obtain reconstructed data. During the training process, unsupervised learning is performed on the similarity between the final training data and the reconstructed data. A first high-dimensional feature is obtained based on the final training data, and a second high-dimensional feature is obtained based on the reconstructed data; Based on the loss function of the PolSAR target detection network, the difference between the final training data and the reconstructed data is minimized, and the difference between the first high-dimensional feature and the second high-dimensional feature is also minimized, thus obtaining the trained PolSAR target detection network.
3. The PolSAR target detection method based on polarization characteristics and anomaly detection according to claim 2, characterized in that, The encoder includes encoder1, encoder2, encoder3, encoder4, and encoder5 modules. The input data size of the encoder is H×W×K, where H is the height, W is the width, and K is the channel dimension. The encoder1 module includes a convolutional layer L1 and a convolutional layer L2. The convolutional layer L1 is used to output a1 feature maps of size H×W. The convolutional layer L1 outputs a1 feature maps The input is fed into the convolutional layer L2, which outputs a1 feature maps of size H×W. The convolutional layer L2 outputs a1 feature maps Input is sent to the encoder2 module; The encoder2 module includes a pooling layer P1, a convolutional layer L3, and a convolutional layer L4. The pooling layer P1 is used to output a1 units. Size feature map The pooling layer P1 outputs a1 feature maps The input is fed into the convolutional layer L3, which is used to output a2 values. Size feature map The convolutional layer L3 outputs a2 feature maps The input is fed into the convolutional layer L4, which is used to output a2 values. Size feature map The convolutional layer L4 outputs a2 feature maps The input is fed into the encoder3 module, and the a2 feature maps output by the convolutional layer L4 are the output features C2 of the encoder, where a2 = 2 * a1; The encoder3 module includes a pooling layer P2, a convolutional layer L5, and a convolutional layer L6. The pooling layer P2 is used to output a2 values. Size feature map The pooling layer P2 outputs a2 feature maps The input is fed into the convolutional layer L5, which is used to output a3 values. Size feature map The convolutional layer L5 outputs a3 feature maps. The input is fed into the convolutional layer L6, which is used to output a3 values. Size feature map The L6 convolutional layer outputs a3 feature maps. The input is fed into the encoder4 module, and the L6 convolutional layer outputs a3 feature maps. For the output feature C3 of the encoder, a3 = 2 * a2; The encoder4 module includes a pooling layer P3, a convolutional layer L7, and a convolutional layer L8. The pooling layer P3 is used to output a3 values. Size feature map The pooling layer P3 outputs a3 feature maps The input is fed into the convolutional layer L7, which is used to output a4 values. Size feature map The convolutional layer L7 outputs a4 feature maps The input is fed into the convolutional layer L8, which is used to output a4 values. Size feature map The L8 convolutional layer outputs a4 feature maps. The input is fed into the encoder5 module, and the L8 convolutional layer outputs a4 feature maps. For the output feature C4 of the encoder, a4 = 2 * a3; The encoder5 module includes a pooling layer P4, a convolutional layer L9, and a convolutional layer L1. 10 The pooling layer P4 is used to output a4 times. Size feature map The pooling layer P4 outputs a4 feature maps The input is fed into the convolutional layer L9, which is used to output a4 values. Size feature map The L9 convolutional layer outputs a4 feature maps. Input to the convolutional layer L 10 The convolutional layer L 10 Used to output 4 Size feature map The convolutional layer L 10 Output a4 feature maps C5 is the output feature of the encoder.
4. The PolSAR target detection method based on polarization characteristics and anomaly detection according to claim 3, characterized in that, The memory library module includes a block-based mem1 module, a mem2 module, a mem3 module, and a mem4 module, wherein: The mem1 module takes the output feature C2 of the encoder as input. First, it performs a block operation on the output feature C2 in the spatial dimension, dividing it evenly into a1 blocks of equal size. Each block is then flattened into a vector, resulting in a1 query vectors. For each query vector Calculate its relationship with the row vector in the memory library M1. The similarity between them yields a set of attention weights. And reset the attention weights below the preset threshold to 0, and then use this set of attention weights. The row vectors in the memory library M1 are weighted, and the weighted vectors are summed to obtain a new feature vector. Therefore, based on the above a1 query vectors... We obtain a1 new eigenvectors Finally, the above new feature vectors are... Reassembling yields a new set of feature maps of the same size as C2. This set of feature maps This refers to the output feature C2′ of the mem1 module; The mem2 module takes the output feature C3 of the encoder as input. First, it performs a block operation on the output feature C3 in the spatial dimension, dividing it evenly into a1 blocks of equal size. Each block is then flattened into a vector, resulting in a1 query vectors. For each query vector Calculate its relationship with the row vector in the memory library M2. The similarity between them yields a set of attention weights. And reset the attention weights below the preset threshold to 0, and then use this set of attention weights. The row vectors in the memory library M2 are weighted, and the weighted vectors are summed to obtain a new feature vector. Based on the above a1 query vectors... We obtain a1 new eigenvectors Finally, the above new feature vectors are... Reassembling yields a new set of feature maps of the same size as C3. This set of feature maps This refers to the output feature C′3 of the mem2 module; The mem3 module takes the output feature C4 of the encoder as input. First, it performs a block operation on the output feature C4 in the spatial dimension, dividing it evenly into a1 blocks of equal size. Each block is then flattened into a vector, resulting in a1 query vectors. For each query vector Calculate its relationship with the row vector in the memory library M3. The similarity between them yields a set of attention weights. And reset the attention weights below the preset threshold to 0, and then use this set of attention weights. The row vectors in the memory library M3 are weighted, and the weighted vectors are summed to obtain a new feature vector. Therefore, based on the above a1 query vectors... We obtain a1 new eigenvectors Finally, the above new feature vectors are... Reassembling yields a new set of feature maps of the same size as C4. This set of feature maps That is, the output feature C′4 of the mem3 module; The mem4 module takes the output feature C5 of the encoder as input. First, it performs a block operation on the output feature C5 in the spatial dimension, dividing it evenly into a1 blocks of equal size. Each block is then flattened into a vector, resulting in a1 query vectors. For each query vector Calculate its relationship with the row vector in the memory library M4. The similarity between them yields a set of attention weights. And reset the attention weights below the preset threshold to 0, and then use this set of attention weights. The row vectors in the memory library M4 are weighted, and the weighted vectors are summed to obtain a new feature vector. Therefore, based on the above a1 query vectors... We obtain a1 new eigenvectors Finally, the above new feature vectors are... Reassembling yields a new set of feature maps of the same size as C5. This set of feature maps This is the output feature C′5 of the mem4 module.
5. The PolSAR target detection method based on polarization characteristics and anomaly detection according to claim 4, characterized in that, The decoder includes decoder1, decoder2, decoder3, decoder4, and decoder5 modules, wherein: The decoder5 module includes an upsampling layer U1 and a convolutional layer L. 11 and convolutional layer L 12 The input of the upsampling layer U1 is the output feature C5′. The upsampling layer U1 first performs bilinear interpolation on the output feature C5′ to obtain a4 values. Size feature map Then, the above four feature maps a4 are analyzed. Perform a convolution operation to obtain a4 elements. Size feature map The a4 feature maps output by the upsampling layer U1 Concatenated with the output feature C′4, a5 are obtained. Size feature map The spliced a5 feature maps Input to the convolutional layer L 11 The convolutional layer L 11 Used to output 4 Size feature map The convolutional layer L 11 Output a4 feature maps Input to the convolutional layer L 12 The convolutional layer L 12 Used to output 4 Size feature map a5 = 2 * a4; The decoder4 module includes an upsampling layer U2 and a convolutional layer L. 13 and convolutional layer L 14 The input of the upsampling layer U2 is the input of the convolutional layer L. 12 The output feature map, the upsampling layer U2 first performs bilinear interpolation on the feature map to obtain a4 features. Size feature map Then, the above four feature maps a4 are analyzed. Perform a convolution operation to obtain a3 values. Size feature map The a3 feature maps output by the upsampling layer U2 Concatenated with the output feature C3′, we obtain a4. Size feature map The stitched a4 feature maps Input to the convolutional layer L 13 The convolutional layer L 13 Used to output a3 Size feature map The convolutional layer L 13 Output a3 feature maps Input to the convolutional layer L 14 The convolutional layer L 14 Used to output a3 Size feature map The decoder3 module includes an upsampling layer U3 and a convolutional layer L. 15 and convolutional layer L 16 The input of the upsampling layer U3 is the input of the convolutional layer L. 14 The output feature map is first subjected to bilinear interpolation by the upsampling layer U3 to obtain a3 features. Size feature map Then, the above three feature maps a3 are analyzed. Perform a convolution operation to obtain a2 values. Size feature map The a2 feature maps output by the upsampling layer U3 Concatenated with the output feature C2′, a3 are obtained. Size feature map The spliced a3 feature maps Input to the convolutional layer L 15 The convolutional layer L 15 Used to output a2 Size feature map The convolutional layer L 15 Output a2 feature maps Input to the convolutional layer L 16 The convolutional layer L 16 Used to output a2 Size feature map The decoder2 module includes an upsampling layer U4 and a convolutional layer L. 17 The input of the upsampling layer U4 is the L 16 The output feature map is first subjected to bilinear interpolation by the upsampling layer U4 to obtain a2 H×W feature maps. Then, the above a2 feature maps Perform convolution operations to obtain a1 feature maps of size H×W. The upsampling layer U4 outputs a1 feature maps Input to the convolutional layer L 17 The convolutional layer L 17 Used to output a1 feature maps of size H×W The decoder1 module includes a convolutional layer L. 18 The L 17 The output feature map is input into the convolutional layer L. 18 The convolutional layer L 18 Used to output 6 H×W feature maps The L 18 Output a6 feature maps As reconstructed data.
6. The PolSAR target detection method based on polarization characteristics and anomaly detection according to claim 2, characterized in that, The PolSAR target detection network also includes a discriminator, wherein: The final training data is input into the discriminator to obtain the first high-dimensional feature; The reconstructed data is input into the discriminator to obtain the second high-dimensional feature.
7. The PolSAR target detection method based on polarization characteristics and anomaly detection according to claim 6, characterized in that, The discriminator includes modules conv1, conv2, conv3, conv4, conv5, and conv6, wherein: The conv1 module includes convolutional layers L 19 The data to be judged is input into the convolutional layer L. 19 The convolutional layer L 19 Used to output a1 Size feature map The data to be discriminated includes the final training data or the reconstructed data; The conv2 module includes convolutional layers L 20 The convolutional layer L 19 The output feature map is input into the convolutional layer L. 20 The convolutional layer L 20 Used to output a2 Size feature map The conv3 module includes convolutional layers L 21 The convolutional layer L 20 The output feature map is input into the convolutional layer L. 21 The convolutional layer L 21 Used to output a3 Size feature map The conv4 module includes convolutional layers L 22 The convolutional layer L 21 The output feature map is input into the convolutional layer L. 22 The convolutional layer L 22 Used to output 4 Size feature map The conv5 module includes convolutional layers L 23 The convolutional layer L 22 The output feature map is input into the convolutional layer L. 23 The convolutional layer L 23 Used to output a5 Size feature map The conv6 module includes convolutional layers L 24 The convolutional layer L 23 The output feature map is input into the convolutional layer L. 24 The convolutional layer L 24 Used to output a2 Size feature map The convolutional layer L 24 The output a2 feature maps are used as high-dimensional features.
8. The PolSAR target detection method based on polarization characteristics and anomaly detection according to claim 2, characterized in that, The loss functions include the L1 Loss function, the SSIM Loss function, and the MSE Loss function, where: The L1 Loss function is expressed as: Where L_L1 represents the L1 Loss function, P represents the set of pixel coordinates, p represents the point coordinates, x represents the final training data, y represents the reconstructed data, and N represents the total number of pixels; The SSIM Loss function is expressed as follows: Where L_SSIM represents the SSIM Loss function, μ x and Let μ represent the mean and variance of x, respectively. y and Let σ represent the mean and variance of y, respectively. xy C1 and C2 are constants used to avoid instability when the SSIM value is 0. The MSE Loss function is expressed as follows: Where L_MSE represents the MSE Loss function, P represents the set of pixel coordinates of the feature map in the high-dimensional feature, and p represents the pixel coordinates of the feature map in the high-dimensional feature.
9. The PolSAR target detection method based on polarization characteristics and anomaly detection according to claim 1, characterized in that, The anomaly detection score map is mapped back to the PolSAR image to be detected, and segmented according to a threshold to obtain the final binary detection result, including: The anomaly detection score map of each test slice is mapped back to the location of the PolSAR image to be tested. For the non-overlapping area of two adjacent test slices, the score of the corresponding anomaly detection score map is taken. For the overlapping area of two adjacent test slices, the score of each pixel in the overlapping area is taken as the larger value of the scores in the anomaly detection score maps corresponding to the two adjacent test slices, thereby obtaining the final anomaly detection score map of the PolSAR image to be tested. The relationship between the final score map of anomaly detection and the threshold th is determined. If the score in the final score map of anomaly detection is greater than the threshold th, the value of the corresponding pixel in the PolSAR image to be tested is set to 1. If the score in the final score map of anomaly detection is less than or equal to the threshold th, the value of the corresponding pixel in the PolSAR image to be tested is set to 0, thus obtaining the final binary detection result.
Citation Information
Patent Citations
A CFAR-guided PolSAR target detection method based on polarimetric characteristics and superpixel segmentation
CN116958676B
POLSAR image marine target detection method based on polarization direction angle compensation
CN104318572A
Polarized synthetic aperture radar image ship detection method based on convolutional neural network
CN113643284A
