SAR image change detection method combining convolution and mixed attention
By combining convolution and hybrid attention in SAR image change detection, and utilizing the composite neighborhood intensity difference method and hierarchical FCM clustering algorithm, a hybrid attention neural network is constructed. This solves the problem of extracting and fusing local details and global information in remote sensing images, and improves the accuracy and robustness of remote sensing image change detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHEJIANG UNIV OF TECH
- Filing Date
- 2026-02-04
- Publication Date
- 2026-05-05
AI Technical Summary
Existing technologies struggle to effectively capture local and global details, especially in boundary regions, in remote sensing image change detection, resulting in poor detection performance.
A SAR image change detection method combining convolution and hybrid attention is proposed. The difference map is analyzed by the composite neighborhood intensity difference method, and the hierarchical FCM clustering algorithm is used for pre-classification. A neural network combining convolution and hybrid attention is constructed, and self-attention and cross-boundary attention are fused for feature extraction and classification.
It improves the accuracy and robustness of remote sensing image change detection, effectively suppresses noise interference, and enhances the ability to extract and process boundary features.
Smart Images
Figure CN121982575A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of synthetic aperture radar image detection methods, specifically a SAR image change detection method that combines convolution and hybrid attention. Background Technology
[0002] With the widespread application of satellite imaging technology in daily life, remote sensing image data of various terrains are becoming increasingly abundant. In the fields of computer vision and intelligent recognition of remote sensing images, change detection in remote sensing images has become an important direction. Although most related research focuses on the overall feature extraction and processing of the change detection area, the extraction and analysis of boundary features are relatively insufficient. In fact, the boundaries of remotely sensed ground features often contain rich semantic information. Therefore, the extraction and processing of boundary information is of great significance for improving the accuracy of change detection in remote sensing images.
[0003] In recent years, deep learning techniques, especially convolutional neural networks (CNNs), have been widely applied to change detection tasks. Typical implementations include dual-stream encoder-decoder structures, Siamese networks, and U-Net-based architectures. These methods can automatically extract multi-level spatial features and improve detection performance to some extent. However, because CNNs primarily focus on information processing within the local receptive field, their ability to model long-range dependencies is limited, making it difficult to effectively capture changes in large-scale ground features or global semantic consistency in remote sensing images.
[0004] To overcome these limitations, researchers began exploring the introduction of Vision Transformer (ViT) into change detection tasks. ViT utilizes a self-attention mechanism to model global interactions between image patches, significantly enhancing the model's ability to understand contextual information. However, when processing high-resolution remote sensing images, ViT faces significant computational overhead and performs poorly in capturing local details, especially in boundary regions, where actual ground changes often first manifest themselves at these edges or contours. Therefore, effectively extracting and fusing local and global information has become a key challenge for change detection performance. Summary of the Invention
[0005] This invention provides a SAR image change detection method that combines convolution and hybrid attention to solve the problems in the background art.
[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0007] A SAR image change detection method combining convolution and hybrid attention includes the following steps:
[0008] Step 1: Acquire two SAR images at different times;
[0009] Step 2: Perform composite neighborhood intensity difference analysis on the two SAR images obtained in Step 1 to obtain a difference map;
[0010] Step 3: Perform pre-classification processing on the difference images obtained in Step 2, dividing the difference images into three categories: changed, unchanged, and uncertain. Select a certain proportion of sample points from the changed and unchanged categories to construct the training dataset, and construct the test dataset with all pixels of the two SAR images.
[0011] Step 4: Construct a neural network that combines convolution and hybrid attention;
[0012] Step 5: Use the training dataset obtained in Step 3 to train the neural network combining convolution and hybrid attention constructed in Step 4;
[0013] Step 6: Use the trained neural network combining convolution and hybrid attention to test the test dataset, obtain the predicted labels of the test dataset, and then obtain the change detection results of the entire image.
[0014] Furthermore, step 1 also includes: performing radiometric correction and geometric registration preprocessing on the original SAR image to obtain SAR images at two different times.
[0015] In the further step 2, the composite neighborhood intensity difference method used in step 2 fuses the neighborhood difference operator and the relative intensity difference in a texture adaptive weighting mechanism to generate a smoother difference map.
[0016] Furthermore, the neural network combining convolution and hybrid attention constructed in step 4 consists of a convolution module and a hybrid attention module, while the hybrid attention is composed of a self-attention and cross-boundary attention fusion module through adaptive weight learning.
[0017] Furthermore, in step 5, the self-attention module divides the training sample set obtained in step 3 into non-overlapping groups. Image patch The value is 3; a linear embedding layer is applied to each image patch, converting it into a fixed-dimensional token embedding; the embedded tokens are then linearly transformed to compute the query. ,key Sum The three input features obtained by linear transformation are input into the self-attention calculation to obtain the self-attention output sequence; the self-attention output sequence is converted into output features through the output projection layer.
[0018] Furthermore, in step 5, the cross-boundary attention is composed of boundary attention and cross attention fused by an adaptive weight learning module; the boundary attention uses two depthwise separable convolutional layers and performs differentiable edge detection to obtain an edge intensity map; the edge intensity map is normalized to obtain an edge spatial weight map; a lightweight channel attention is used to enhance the adaptability of the channel dimension; the channel attention weight map and the edge spatial weight map are multiplied element-wise to obtain the boundary weights; the boundary weights are multiplied element-wise with the input features to obtain the boundary enhancement features; the boundary enhancement features are flattened into a spatial sequence form, using the first SAR image as an example. The boundary enhancement feature is With the second SAR image The boundary enhancement feature is and Calculate multi-head attention; linearly project the attention output to reshape it into a spatial tensor, and then... Features are fused through convolutional layers, batch normalization layers, and activation functions to obtain enhanced first-phase features. Second phase characteristics Remain unchanged; combine the enhanced features with the second phase features. By performing channel-dimensional connections, enhanced cross-boundary attention features are obtained.
[0019] Furthermore, in step 5, a concatenation operation is used to connect the global features output by self-attention and the enhanced cross-boundary features obtained by cross-boundary attention in the channel dimension; a fusion weight matrix is generated using a convolutional layer and a sigmoid activation function; the global features and the enhanced cross-boundary features are weighted and fused according to the weight matrix; and the fused features are obtained through learnable residual connections.
[0020] Furthermore, after performing a concatenation operation on the channel dimension on the fused features and local features, the data is classified through three fully connected networks to obtain the predicted labels of change and invariance. The loss function of the model is calculated based on the predicted label results, and the Adam optimizer is used to optimize the parameters of the neural network that combines convolution and hybrid attention.
[0021] Furthermore, the formula for calculating the channel attention weight map is as follows: ;
[0022] in, Indicates input features, Indicates global average pooling. This represents the activation function. and Indicates a fully connected layer. express Activation function.
[0023] Furthermore, in step 5, the convolution module includes using a large convolution kernel to perform feature transformation and channel compression on the input features, followed by normalization and... After activation, local features are obtained; a convolutional kernel of the same size is used to restore the number of channels to the original dimension, and normalization is applied again. The activation function yields enhanced local features.
[0024] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0025] 1) This invention employs a composite neighborhood intensity difference method for difference map analysis to obtain difference maps between two SAR images. The composite neighborhood intensity difference method synergistically combines neighborhood structural differences and pixel-level relative intensity changes, and balances the importance of different features through an adaptive mechanism, thereby enhancing the variability of the samples.
[0026] 2) This invention uses hierarchical FCM (fuzzy C-means) clustering algorithm for pre-classification to obtain a pseudo-label matrix. Hierarchical FCM clustering has good results and high efficiency, thus improving the accuracy and speed of pre-classification.
[0027] 3) This invention effectively combines convolution with hybrid attention, thereby making the change detection results more robust, the effect better, and the classification accuracy of the network more precise. Attached Figure Description
[0028] Figure 1 This is a flowchart of the method in Embodiment 1 of the present invention.
[0029] Figure 2 This is a schematic diagram of the neural network structure described in Embodiment 1 of the present invention;
[0030] Figure 3 These are the change detection results on SAR images in the Yellow River region during the experiment of Embodiment 2 of the present invention, wherein: (a) is a SAR image of the Yellow River during the flood season, (b) is a SAR image of the Yellow River during the dry season, (c) is a manually interpreted reference image, and (d)~(h) correspond to the change detection results of the adaptive structure graph method, log-likelihood ratio method, similarity graph regression method, neighborhood difference method and the method of the present invention, respectively.
[0031] Figure 4 These are the change detection results on SAR images in Lu'an City during the experiment of Embodiment 2 of the present invention, wherein: (a) is a SAR image of Lu'an City during the flood period, (b) is a normal SAR image of Lu'an City, (c) is a manually interpreted reference image, and (d)~(h) correspond to the change detection results of the adaptive structure graph method, log-likelihood ratio method, similarity graph regression method, neighborhood difference method and the method of the present invention, respectively. Detailed Implementation
[0032] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0033] Example 1
[0034] like Figure 1 As shown, this embodiment discloses a SAR image change detection method combining convolution and hybrid attention, including the following steps:
[0035] Step 1: Acquire raw SAR images from synthetic aperture radar. and Radiometric correction and geometric registration preprocessing are performed on the original SAR images.
[0036] Step 2: Obtain the two SAR images obtained in Step 1. and A difference map is generated using the composite neighborhood intensity difference method. The formula for calculating the difference map is: ( =1,2); ; ; ; ; ; ; ;
[0037] in, Indicates the number of images. The mean of two SAR images is given. These represent two SAR images after denoising. Indicates use Mean filtering; For local variance, The local standard deviations of the two SAR images; Represents a texture map; Representing the neighborhood operator, (t=1,2) represents the denoised SAR image at the t=1,2th t-th t- Each component ( =1, ...,h2); For intensity difference operators, It is a parameter to prevent the denominator from being zero; This is a difference map between two SAR images. This represents the difference plot after min-max normalization. Adaptive weights.
[0038] Step 3: Compare the differences. Perform pre-classification and construct training and test datasets:
[0039] (3.1) Pre-classification: For difference plots The hierarchical FCM (fuzzy C-means) clustering algorithm is used for pre-classification to obtain a pseudo-label matrix with pseudo-label values of 0, 1, and 0.5. This means that after pre-classification, high-probability samples that are judged as unchanged or changed, as well as uncertain samples, are generated.
[0040] (3.2) Training sample data block generation: Select 5-15% of the pixels with pseudo-label values of 0 and 1 as pseudo-labels and extract the spatial locations of these pixels. On the corresponding spatial locations of the pixels in the two original SAR images, take the surrounding area of the pixel. The neighboring pixel blocks are used as the training set, and the edge pixels are extracted by padding the neighboring pixels with 0. The value of is an odd number that is not less than 3;
[0041] (3.3) Test sample data block generation: Extract all pixels around the original two SAR images The neighboring pixel blocks are used as the test set, and the edge pixels are extracted by padding the neighboring pixels with 0. The value of is an odd number that is not less than 3.
[0042] Step 4: Construct a convolutional and hybrid attention neural network:
[0043] The constructed neural network combining convolution and hybrid attention consists of convolutional modules and hybrid attention modules, while the hybrid attention consists of self-attention and cross-boundary attention.
[0044] (4.1) Self-attention module:
[0045] The two SAR images obtained in step 1 are divided into non-overlapping segments. Image patches are processed, and each image patch is encoded into a token embedding through a linear embedding layer; the embedded tokens are then subjected to a linear transformation to compute the query. ,key Sum The three input features obtained by linear transformation are input into the self-attention calculation to obtain the self-attention output sequence; the self-attention output sequence is converted into output features through the output projection layer.
[0046] Self-attention output is calculated using the following formula: ;
[0047] in, For self-attention output, This is the scaling factor.
[0048] (4.2) Cross-boundary attention:
[0049] The constructed cross-boundary attention consists of boundary attention and cross attention.
[0050] (4.2.1) Construct the boundary attention module:
[0051] The boundary attention module consists of two depthwise separable convolutional layers. and Perform Sobel convolution operations in the X and Y directions respectively, and then obtain the edge intensity map. Global normalization is then performed to obtain the edge space weight map. To further enhance the adaptability of the channel dimension, a lightweight channel attention mechanism is used. Final boundary weights for: ;
[0052] in, Input features for two SAR images, For a very small parameter, Indicates global average pooling. This represents the activation function. and Indicates a fully connected layer. express Activation function This indicates element-wise multiplication.
[0053] (4.2.2) Constructing a cross-boundary attention module
[0054] The two SAR images obtained in step 1 are divided into... The image blocks are processed by filling the edge pixels with zeros. A boundary attention module is applied to each image to generate corresponding boundary weights. The boundary weights of the two SAR images are then obtained using the boundary attention module. and Then respectively and Element-wise multiplication with the input features yields the boundary enhancement features. and Subsequently, the two were flattened into a spatial sequence and displayed using the first SAR image. The boundary enhancement feature is With the second SAR image The boundary enhancement feature is and Compute multi-head attention to construct a cross-temporal attention mechanism; linearly project the attention output to reshape it into a spatial tensor, and then... Features are fused through convolutional layers, batch normalization layers, and activation functions to obtain enhanced first-phase features. Second phase characteristics Remain unchanged; combine the enhanced features with the second phase features. Channel-level connections are performed to obtain enhanced cross-boundary attention features; the calculation formula is as follows: ; ; ; ; ;
[0055] in, This represents the boundary weights generated by the boundary attention module for the two SAR images. It is an edge weight graph, where, and It is to utilize The operator performs differentiable edge detection using two depthwise separable convolutional layers. It is a parameter for stable numerical stability; The input features representing dual-temporal SAR images, Represents the boundary enhancement features of dual-temporal images; It is a boundary correlation matrix generated by the outer product of two boundary maps, which serves as an attention bias term to emphasize semantic matching between high-gradient regions. Indicates the outer product operation; coefficient =0.1 is used to balance self-attention and boundaries; It is a lightweight attention weight.
[0056] (4.3) Construction of the adaptive weight learning fusion module:
[0057] The global features output by the self-attention method and the enhanced cross-boundary features obtained by the cross-boundary attention method are connected in the channel dimension using a concatenation operation; a fusion weight matrix is generated using a convolutional layer and a sigmoid activation function; the global features and the enhanced cross-boundary features are weighted and fused according to the weight matrix; and the fused features are obtained through learnable residual connections.
[0058] (4.4) After performing a concatenation operation on the channel dimension on the fused features and local features, the network is classified through three fully connected networks to obtain the predicted label results of change and invariance; the loss function of the model is calculated based on the predicted label results, and the Adam optimizer is used to optimize the parameters of the neural network that combines convolution and hybrid attention.
[0059] (4.5) Convolutional module:
[0060] The training sample set obtained in step 3 is first processed by a large convolutional kernel to perform feature transformation and channel compression on the input features, followed by normalization and... After activation, local features are obtained; then, a convolutional kernel of the same size is used to restore the number of channels to the original dimensions, and normalization is applied again. The activation function preserves details and yields enhanced local features. The calculation results are as follows: ; ;
[0061] in, These are intermediate variables after dimensional compression. This represents the extracted enhanced local features; Indicates the kernel size as ( =7), This indicates a batch normalization operation. This represents the activation function.
[0062] Step 5: Use the training set obtained in Step 3 to train the neural network that combines convolution and hybrid attention.
[0063] Step 6: Use the trained neural network combining convolution and hybrid attention to test the test dataset, obtain the predicted labels of the test dataset, and then generate the final change detection result map.
[0064] Example 2
[0065] This embodiment is an experimental example. Two sets of real SAR image data were selected for comparative experiments to further illustrate the effectiveness of the change detection method described in Embodiment 1.
[0066] 1. Experimental setup:
[0067] Experimental environment: Python 3.10.12, PyTorch 2.1.0, Intel(R) Core(TM) i7-14650HX processor, NVIDIA RTX 4060 graphics card, Windows 11.
[0068] Experimental data: Figure 3 Images of the Yellow River dataset, taken by the Radarset-2 synthetic aperture radar sensor in June 2008 and June 2009, covering the Yellow River estuary region within China, with a spatial resolution of 8m and an image size of 257×289 pixels. Figure 4 The images are SAR images of Lu'an City before and after the flood disaster, acquired by the Sentinel-1 SAR satellite. The spatial resolution is 10m and the image size is 429×580 pixels.
[0069] Experimental parameters: Sample image patch size =9, non-overlapping image patches =3.
[0070] Experimental comparison methods: Adaptive structure graph method, log-likelihood ratio method, similarity graph regression method, and neighborhood difference method were used respectively.
[0071] Experimental evaluation indicators: The number of missed detections, the number of false detections, the accuracy rate, and the Kappa coefficient were used to quantitatively evaluate the change detection results.
[0072] 2. Experimental Results and Analysis:
[0073] Experiment 1 applied the adaptive structure graph method, log-likelihood ratio method, similarity graph regression method, neighborhood difference method, and the method of this invention to detect changes in SAR images of the Yellow River during the flood and drought seasons, respectively. The results are as follows: Figure 3 As shown, Figure 3 (a) is a SAR image of the Yellow River during the flood season. Figure 3 (b) is a SAR image of the Yellow River during the dry season. Figure 3 (c) is a reference image for manual interpretation. Figure 3 (d)~(h) correspond to the change detection results of the adaptive structure graph method, log-likelihood ratio method, similarity graph regression method, neighborhood difference method and the method of the present invention, respectively.
[0074] Experiment 2 employed four existing change detection methods in conjunction with the method of this invention to perform change detection on SAR images near Lu'an City. The results are as follows: Figure 4 As shown, where Figure 4 (a) is a SAR image of the Lu'an City area during the flooding period. Figure 4 (b) is a regular SAR image of Lu'an City. Figure 4 (c) is a reference image for manual interpretation. Figure 4(d)~(h) correspond to the change detection results of the adaptive structure graph method, log-likelihood ratio method, similarity graph regression method, neighborhood difference method and the method of the present invention, respectively.
[0075] from Figure 3 , Figure 4 As shown in Table 1, the overall detection accuracy of the change detection method described in Example 1 is superior to other change detection methods. This is because the change detection method in Example 1 uses cross-boundary attention. Due to the influence of speckle noise on SAR images, the boundaries of changed regions are often blurred. These regions often contain weak but crucial change signals. Cross-boundary attention aligns the boundary features of the cross-SAR images, enhancing the contrast between changed and unchanged regions and reducing noise interference in change discrimination. Furthermore, self-attention and cross-boundary attention interact and are fused with local feature information from the convolutional part, enhancing feature diversity, helping to suppress SAR speckle noise, reduce false alarm rate, and resulting in more robust and accurate change detection results. Table 1 is shown below:
[0076] Table 1. SAR image change detection results in the Yellow River and Lu'an City areas
[0077] The experimental results above demonstrate that the present invention has clear advantages in noise resistance and detection accuracy compared with existing SAR image change detection methods.
[0078] The embodiments described herein are merely preferred embodiments of the invention and are not intended to limit the concept and scope of the invention. Any modifications and improvements made by those skilled in the art to the technical solutions of the invention without departing from the design concept of the invention should fall within the protection scope of the invention. The technical content for which protection is sought in this invention has been fully described in the claims.
Claims
1. A SAR image change detection method combining convolution and hybrid attention, characterized in that, Includes the following steps: Step 1: Obtain two SAR images of the same geographic area taken at different times; Step 2: Perform composite neighborhood intensity difference analysis on the two SAR images obtained in Step 1 to generate difference images; Step 3: Perform pre-classification processing on the difference images obtained in Step 2, dividing the difference images into three categories: changed, unchanged, and uncertain. Select a certain proportion of sample points from the changed and unchanged categories to construct the training dataset, and construct the test dataset with all pixels of the two SAR images. Step 4: Construct a neural network that combines convolution and hybrid attention; The constructed neural network combining convolution and hybrid attention consists of a convolution module and a hybrid attention module. The hybrid attention module is composed of a self-attention and cross-boundary attention module that are fused through adaptive weight learning. Step 5: Use the training dataset obtained in Step 3 to train the neural network combining convolution and hybrid attention constructed in Step 4; Step 6: Use the trained neural network combining convolution and hybrid attention to test the test dataset, obtain the predicted labels of the test dataset, and then generate the final change detection result map.
2. The SAR image change detection method combining convolution and hybrid attention as described in claim 1, characterized in that, In step 1, radiometric correction and geometric registration preprocessing are performed on the original SAR image to obtain two SAR images at different times.
3. The SAR image change detection method combining convolution and hybrid attention as described in claim 1, characterized in that, The composite neighborhood intensity difference method used in step 2 fuses the neighborhood difference operator and the relative intensity difference in a texture-adaptive weighting mechanism to generate a smoother difference map.
4. The SAR image change detection method combining convolution and hybrid attention as described in claim 1, characterized in that, Step 3 specifically includes the following steps: (3.1) The hierarchical FCM clustering algorithm is used to preclassify the difference map to obtain a pseudo-label matrix with pseudo-label values of 0, 1 and 0.5 respectively. That is, after preclassification, high-probability samples that are judged as unchanged or changed and uncertain samples are generated. (3.2) Select 5-15% of the pixels with pseudo-label values of 0 and 1, and based on the spatial location of these pixels, extract the surrounding pixels from the original two SAR images. Image patches were used as the training set. Edge pixels were extracted by padding their neighborhoods with zeros. The value of is an odd number that is not less than 3; (3.3) Extract the surrounding area of all pixels from the original two SAR images. Image patches were used as the test set. Edge pixels were extracted by padding their neighborhoods with zeros. The value of is an odd number that is not less than 3.
5. The SAR image change detection method combining convolution and hybrid attention as described in claim 1, characterized in that, Step 4, constructing a neural network combining convolution and hybrid attention, is specifically implemented in the following way: (4.1) Construction of self-attention module: The training sample set obtained in step 3 is divided into non-overlapping groups. Image patch The value is 3; a linear embedding layer is applied to each image patch, converting it into a fixed-dimensional token embedding; the embedded tokens are then linearly transformed to compute the query. ,key Sum The three input features obtained from the linear transformation are input into the self-attention calculation to obtain the self-attention output sequence; the self-attention output sequence is then converted into output features through the output projection layer. (4.2) Construction of the cross-boundary attention module: The cross-boundary attention consists of a boundary attention module and a cross-attention module fused by adaptive weight learning; the two SAR images obtained in step 1 are divided into... Image blocks, The value is 9; a boundary attention module is applied to each image to generate corresponding boundary weights. The boundary attention module uses two depthwise separable convolutional layers and performs differentiable edge detection to obtain an edge intensity map; the edge intensity map is normalized to obtain an edge space weight map; a lightweight channel attention is used to enhance the adaptability of the channel dimensions; the channel attention weight map and the edge space weight map are multiplied element-wise to obtain the boundary weights; the boundary weights are multiplied element-wise with the input features to obtain the boundary enhancement features; Flatten the boundary enhancement features into a spatial sequence form, using the first SAR image. The boundary enhancement feature is With the second SAR image The boundary enhancement feature is and Calculate multi-head attention; linearly project the attention output to reshape it into a spatial tensor, and then... Features are fused using convolutional layers, batch normalization layers, and activation functions; The enhanced first phase features were obtained. Second phase characteristics Remain unchanged; combine the enhanced features with the second phase features. By performing channel-dimensional connections, enhanced cross-boundary attention features are obtained; (4.3) Construction of the adaptive weight learning fusion module: The global features output by the self-attention and the enhanced cross-boundary features obtained by the cross-boundary attention are connected in the channel dimension using a concatenation operation; a fusion weight matrix is generated using a convolutional layer and a sigmoid activation function; the global features and the enhanced cross-boundary features are weighted and fused according to the weight matrix; and the fused features are obtained through learnable residual connections. (4.4) After performing a concatenation operation on the channel dimension on the fused features and local features, the network is classified through three fully connected networks to obtain the predicted label results of change and invariance; the loss function of the model is calculated based on the predicted label results, and the Adam optimizer is used to optimize the parameters of the neural network that combines convolution and hybrid attention.
6. The SAR image change detection method combining convolution and hybrid attention as described in claim 1, characterized in that, The formula for calculating the channel attention weight map is as follows: ； in, Indicates input features, Indicates global average pooling. This represents the activation function. and Indicates a fully connected layer. express Activation function.
7. The SAR image change detection method combining convolution and hybrid attention as described in claim 1, characterized in that, The convolution module uses a large convolutional kernel to perform feature transformation and channel compression on the input features, followed by normalization and... After activation, local features are obtained; Using a convolutional kernel of the same size, the number of channels is restored to the original dimension, and normalization is applied again. The activation function yields enhanced local features.