A method for constructing SAR image difference maps based on NSST fusion

Abstract

This invention discloses a SAR image difference map construction method based on NSST fusion, applicable to the field of SAR image transform detection. The method includes: acquiring two SAR images of the same area at different times; constructing a mean difference map and a Gaussian difference map using a mean operator and a Gaussian operator respectively; performing a non-subsampled shear wave NSST transform on the mean difference map and the Gaussian difference map to obtain high- and low-frequency subband coefficients; performing weighted fusion of the low-frequency coefficients based on threshold-based visual saliency mapping, and performing improved spatial frequency maximization fusion of the high-frequency coefficients based on threshold-based thresholding; and performing an inverse NSST transform on the fused low-frequency and high-frequency coefficients to obtain the final fused difference map. This invention fully utilizes the advantages of different difference map operators, preserving image intensity and edge detail information while suppressing background noise, improving the contrast and integrity of the SAR image difference map, and effectively enhancing the accuracy of difference map construction.

Description

Technical Field

[0001] This invention relates to the field of SAR image transformation detection, and specifically to a method for constructing SAR image difference maps based on non-subsampled shearlet transform (NSST) fusion. Background Technology

[0002] Change detection in SAR images is an important branch of SAR image interpretation. It determines information about changes in ground features by analyzing two SAR images of the same area acquired at different times. The difference map-based change detection method is currently recognized as the most effective and widely used SAR image change detection method, and the quality of the difference map directly affects the accuracy of the SAR image change detection results.

[0003] In practical applications, SAR image difference maps can be generated by extracting difference information between SAR images at two different time points using difference operators. Commonly used methods for constructing difference maps include the logarithmic ratio method, the mean operator method, the Gaussian operator method, and the KL divergence method. Because different difference operators measure SAR image changes differently, the difference information extracted by different operators varies. Furthermore, the inherent speckle noise in SAR images increases the difficulty of extracting difference information, making it difficult to directly construct a difference map with low noise levels that best showcases SAR image changes using a single difference operator. Therefore, difference maps from different operators can be complementary and fused, leveraging the advantages of each operator to generate a new fused difference map, thus improving the quality of the generated difference map.

[0004] Existing commonly used multi-operator fusion methods include weighted average-based difference map fusion methods and wavelet transform-based difference map fusion methods. Although these methods can complement and fuse single-operator difference maps to a certain extent, they still have problems such as the fusion rules not fully utilizing image feature information and the fusion process not considering noise suppression. This may lead to phenomena such as missing target area intensity information, blurred edge contours, and poor visual effects in the fusion results. Therefore, researching effective multi-operator fusion methods to fully display the transformation information of SAR images has become the focus of current research. Summary of the Invention

[0005] This invention addresses the problem of constructing SAR image difference maps by providing a method based on NSST fusion. The method fuses mean operator difference maps and Gaussian operator difference maps using NSST transformation. It employs a threshold-based visual saliency map (VSM) and a modified spatial frequency (MSF) index as low / high frequency fusion rules. This approach suppresses noise and preserves edge details during image fusion, improving the accuracy of SAR image difference map construction.

[0006] The SAR image difference map construction method based on NSST fusion of the present invention includes the following steps:

[0007] Step 1: Acquire two SAR images, X1 and X2, of the same area at different times;

[0008] Step 2: Generate the mean difference map X1 and X2 of two SAR images using the mean operator method. A ;

[0009] Step 3: Generate Gaussian difference maps X1 and X2 from two SAR images using the Gaussian operator method. B ;

[0010] Step 4: Analyze the mean difference graph X separately. A Difference diagram X of Gaussian B Perform NSST transformation to obtain high and low frequency subband coefficients; let C AL C BL These are the low-frequency coefficients of the mean difference plot and the Gaussian difference plot, respectively. These are the high-frequency coefficients of the mean difference plot and the Gaussian difference plot in the k-axis at scale l, respectively.

[0011] Step 5: Analyze the low-frequency coefficients C of the mean difference plot and Gaussian difference plot. AL C BL The fusion is performed using a threshold-based VSM weighted fusion rule to obtain the fused low-frequency coefficient C. L ;

[0012] Step Six: Apply the MSF maximization fusion rule based on threshold judgment to the high-frequency coefficients of the mean difference map and Gaussian difference map. The high-frequency coefficients are obtained by fusion.

[0013] Step 7: Combine the fused low-frequency coefficients C L and high frequency coefficient Perform the NSST inverse transform to obtain the final fusion difference map.

[0014] In step five, a VSM weighted fusion rule based on threshold judgment is used to evaluate the low-frequency coefficient C. AL C BL The fusion process includes: first, calculating the low-frequency coefficient image C. AL C BL The normalized visual saliency values ​​at each coordinate (i,j) are calculated; then, the fusion weights are calculated as follows:

[0015]

[0016] Among them, V A (i,j) and V B (i,j) represent the low-frequency coefficient image C, respectively. AL C BL Normalized visual saliency at coordinates (i,j); w A (i,j) and w B (i,j) represent C respectively AL C BL The corresponding fusion weight value;

[0017] Finally, VSM-weighted fusion was performed based on the visual saliency threshold κ to obtain the fused low-frequency coefficients C. L ;

[0018]

[0019] Among them, C L (i,j) represents the low-frequency coefficients fused at coordinates (i,j).

[0020] In step six, a threshold-based MSF maximization fusion rule is used to optimize high-frequency coefficients. The fusion process includes: first, calculating the high-frequency coefficients of the image. The spatial frequency value F at each coordinate (i,j) A (i,j) and F B (i,j); then, based on the spatial frequency threshold ν, the MSF is maximized and fused to obtain the fused high-frequency coefficients.

[0021]

[0022] in, Let (i,j) be the high-frequency coefficients fused at coordinates (i,j). High-frequency coefficient images High-frequency coefficients at coordinates (i,j).

[0023] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0024] The SAR image difference map construction method based on NSST fusion of the present invention first constructs a single-operator difference map using the mean operator with good noise suppression effect and the Gaussian operator with good edge preservation ability. Then, the mean difference map and the Gaussian difference map are complementaryly fused using NSST transformation. The VSM index is introduced as a low-frequency fusion criterion, and the MSF index is used as a high-frequency fusion criterion. Threshold judgment is added in the fusion process. It can preserve the intensity information and edge detail information of the image while effectively suppressing noise, thereby improving the contrast of the SAR image difference map and improving the completeness and accuracy of the difference map construction. Moreover, the core algorithm of the present invention is simple to implement and has high practicality. Attached Figure Description

[0025] Figure 1 This is a flowchart of an implementation of the SAR image difference map construction method based on NSST fusion of the present invention;

[0026] Figure 2 This is a comparison of the results of constructing the Ottawa region difference map using different methods according to embodiments of the present invention;

[0027] Figure 3 This is a comparison of the results of constructing the Yellow River region difference map using different methods in the embodiments of the present invention. Detailed Implementation

[0028] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0029] This invention is a method for constructing SAR image difference maps based on NSST fusion. An implementation process is as follows: Figure 1 As shown, the specific steps include steps one through six.

[0030] Step 1: Acquire two SAR images, X1 and X2, of the same area at different times.

[0031] Step 2: Process SAR images X1 and X2 using the mean operator method, which has a good noise suppression effect, to generate a mean difference map X. A The following is an example:

[0032]

[0033]

[0034] Among them, X A This is a mean difference map, where (i,j) is the coordinate point position in the mean difference map, and μ1(i,j) and μ2(i,j) are the local means of two SAR images X1 and X2 at pixel (i,j). k (i,j) represents the image X kThe gray value at coordinate (i,j) is used to select a local window with a size of 2N+1.

[0035] Step 3: Process SAR images X1 and X2 using the Gaussian operator, which has good edge preservation capabilities, to generate a Gaussian difference map X. B The following is an example:

[0036]

[0037]

[0038]

[0039] in, and Let X1 and X2 be local windows at pixel (i,j) in SAR images X1 and X2, respectively, where the window size is 2N+1, log represents taking the logarithm of the pixels within the local window to compress the gray levels, and G is a rotationally symmetric Gaussian low-pass filter. 1r (i,j) and X 2r (i,j) represent the Gaussian-filtered images of the local windows of SAR images X1 and X2 at pixel (i,j), respectively. * represents absolute value operation, and * represents convolution operation. In this invention, the variance of the Gaussian filter is set to 0.5 based on empirical values. The following is the expression for a 3×3 rotationally symmetric Gaussian low-pass filter:

[0040]

[0041] Where the filter matrix G is the normalization matrix, i.e.

[0042] Step 4: The NSST transform, which has the advantages of multi-scale and multi-directionality, translation invariance and low computational complexity, is used to decompose the mean difference map and Gaussian difference map in turn to obtain the high- and low-frequency subband coefficients. Where C AL C BL The difference between the means is shown in graph X. A Difference diagram X of Gaussian B The low-frequency coefficient, The difference between the means is shown in graph X. A Difference diagram X of Gaussian B High-frequency coefficients in the k-axis at scale l.

[0043] Step 5: Extract the low-frequency component C from the mean difference plot and the Gaussian difference plot. AL C BL A threshold-based VSM weighted fusion strategy is employed to obtain the fused low-frequency component C. LAmong them, the threshold-based VSM fusion strategy can effectively suppress noise, preserve significant visual structure and regional contrast in low-frequency coefficients, highlight the intensity information of the target in the fused image, and thus improve the visual quality of the fused image.

[0044] The visual significance (VSM) of low-frequency coefficients in mean difference plots and Gaussian difference plots is defined as follows:

[0045]

[0046] Where S(i,j) represents the visual saliency of the low-frequency coefficient at the (i,j) coordinate, m and n represent the size of the low-frequency coefficient image L, and L(i,j) and L(q,t) represent the low-frequency coefficient values ​​at the (i,j) and (q,t) coordinates, respectively.

[0047] The visual significance values ​​of the low-frequency coefficients are normalized and then expressed as follows:

[0048]

[0049] Where max(S) and min(S) represent the maximum and minimum visual saliency values ​​of the low-frequency coefficients, respectively.

[0050] Based on the visual significance values ​​of the low-frequency coefficients after VSM processing, the weighted fusion weights of VSM can be set as follows:

[0051]

[0052] Among them, V A (i,j) and V B (i,j) represent the low-frequency coefficients C, respectively. AL C BL The normalized visual saliency value at coordinates (i,j), w A (i,j) and w B (i,j) represent the weight values ​​of the two respectively.

[0053] Building upon traditional weighted fusion rules, this invention adds a threshold judgment based on visual saliency, setting low-frequency coefficients with smaller saliency values ​​to zero. This suppresses noise interference while preserving image overview information. Therefore, the VSM weighted fusion rule based on threshold judgment for low-frequency coefficients is expressed as follows:

[0054]

[0055] Where κ is the visual saliency threshold, and C L C represents the low-frequency coefficients of the fused NSST. L (i,j) represents the low-frequency coefficients fused at coordinates (i,j).

[0056] Step Six: Employ a threshold-based MSF maximization fusion strategy to extract high-frequency components from the mean difference map and Gaussian difference map. The components are fused to obtain the fused high-frequency components. Among them, threshold-based MSF can measure the image sharpness information from the gray-level transformation of the rows, columns, and diagonals of the high-frequency coefficient matrix, thereby improving the texture sharpness and edge contrast of the fused image while suppressing noise.

[0057] The MSF expression for the high-frequency coefficients of the mean difference plot and the Gaussian difference plot is as follows:

[0058]

[0059] Where F(i,j) represents the spatial frequency value of the high-frequency coefficient H at coordinate (i,j), RF and CF represent the row and column frequencies respectively, and DF and LF are the diagonal frequencies. For a window of size M×N centered at the high-frequency coefficient H(i,j), the calculation formulas for RF, CF, DF, and LF are as follows:

[0060]

[0061]

[0062]

[0063]

[0064] This invention adds a spatial frequency-based threshold judgment to the MSF (Maximum Segment Value) fusion strategy, setting high-frequency coefficients with smaller spatial frequency values ​​to zero, thus suppressing noise interference while preserving image detail. The MSF fusion rule based on threshold judgment for high-frequency coefficients is as follows:

[0065]

[0066] Where v is the spatial frequency threshold. For the fused NSST high-frequency coefficients, F A (i,j), F B (i,j) are high-frequency coefficients respectively The corresponding MSF value.

[0067] Step 7: Combine the fused low-frequency component C L and high frequency components Perform the NSST inverse transform to obtain the final fusion difference map.

[0068] Example

[0069] To verify the effectiveness of the method of the present invention in constructing SAR image difference maps, this embodiment compares the commonly used difference map generation methods with the method of the present invention, and uses the signal-to-noise ratio (SNR) and structural similarity (SSIM) as performance evaluation indicators to quantitatively evaluate the difference map construction effect of each method.

[0070] The experiment used multi-temporal SAR images from the Ottawa region and the Yellow River Delta region to generate difference maps. The difference maps were compared with those generated by single operators using logarithmic, mean, and Gaussian operators, as well as those generated by wavelet fusion based on mean and Gaussian operators. The image sizes for the two regions were 290×350 and 288×256 pixels, respectively. All comparison images used were registered.

[0071] Figure 2 Figures (a) and (b) show SAR images acquired by the Radarsat-1 satellite in Ottawa, the capital of Canada, in May and August 1997, respectively. The variations are shown in the reference figure. Figure 2 As shown in (c) Figure 2 (d)-(h) show the differences between SAR images generated using different methods. Figure 2 It can be seen that the logarithmic operator can preserve details such as the edges of the changed region, but its noise suppression effect is not good because it does not introduce spatial information. The mean operator introduces spatial information and has a significant effect on suppressing noise in the invariant region, but its preservation of edge information in the changed region is poor. The Gaussian operator difference map preserves the image edge details well, but its ability to suppress noise in the invariant region is poor, and some isolated noise points exist in the difference map. Although the wavelet fusion method combines the advantages of the Gaussian operator and the mean operator, some noise regions are still preserved in the image. In the difference map of the fusion of the mean operator and the Gaussian operator based on NSST proposed in this invention, the target intensity and details in the image changed region are well preserved. It can be seen that this method enhances the image contrast while also having a good suppression effect on background noise.

[0072] Figure 3 Figures (a) and (b) show SAR images acquired by the Radarsat-2 satellite in the Yellow River Delta region of my country in June 2008 and June 2009, respectively. The changes are shown in the reference figure. Figure 3 As shown in (c) Figure 3 (d)-(h) show the difference maps of SAR images generated using different methods. Figure 3It can be seen that in the Yellow River Delta region, which is severely affected by noise, the regions where the logarithmic operator difference map changes are completely submerged in noise; the mean operator has a good effect on background suppression, but its ability to preserve image details is poor; the Gaussian operator, while preserving image edge information, also enhances noise in the background; although the wavelet fusion operator combines the edge preservation ability of the Gaussian operator and the noise suppression ability of the mean operator, there are still obvious noise regions; the difference map fusion method based on NSST proposed in this invention can suppress noise while preserving the structural information of the transformed region and improving the contrast of the difference map.

[0073] The difference maps generated by the method of this invention and those generated by other comparative methods were quantitatively evaluated, and the evaluation results are shown in Table 1. The signal-to-noise ratio (SNR) evaluation index is used to measure the noise suppression capability of each method, and the structural similarity (SSIM) evaluation index reflects the degree of similarity between the difference maps generated by each method and the transformed reference map. The larger the values ​​of the two evaluation indexes, the better the construction effect of the SAR image difference map.

[0074] Table 1. Evaluation results of the differences between the methods.

[0075]

[0076] As shown in Table 1, the method of this invention achieved the maximum values ​​for the difference map evaluation indices SNR and SSIM in both the Ottawa region and the Yellow River Delta region. Figure 2 and Figure 3 The consistent visual effects demonstrate that the method of the present invention can effectively suppress image noise while improving the preservation of SAR image transformation regions, proving that the method of the present invention is a feasible method for generating SAR image difference maps, which is beneficial for subsequent SAR image change detection.

[0077] Except for the technical features described in the specification, all other technologies are known to those skilled in the art. Descriptions of well-known components and technologies are omitted in this invention to avoid redundancy and unnecessary limitation. The embodiments described above do not represent all embodiments consistent with this application. Various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of this invention are still within the protection scope of this invention.

Claims

1. A method for constructing SAR image difference maps based on NSST fusion, characterized in that, Includes the following steps: Step 1: Acquire two SAR images of the same area at different times. and ; Step 2: Generate two SAR images using the mean operator method. and Mean difference plot ; Step 3: Generate two SAR images using the Gaussian operator method. and Gaussian difference plot ; Step 4: Analyze the mean difference plots separately. Difference plot of Gaussian Perform NSST decomposition to obtain the high- and low-frequency subband coefficients of the mean difference plot and Gaussian difference plot. These are the low-frequency coefficients of the mean difference plot and the Gaussian difference plot, respectively. The mean difference plot and the Gaussian difference plot are respectively in scale High-frequency coefficients in the direction; Step 5: Apply a threshold-based VSM weighted fusion rule to the low-frequency coefficients of the mean difference map and Gaussian difference map. Perform fusion, and assume that the fused low-frequency coefficients are obtained. ; A threshold-based VSM weighted fusion rule is used to evaluate low-frequency coefficients. The fusion process includes the following: First, calculate the visual saliency of the low-frequency coefficients; in the low-frequency coefficient image... Visual saliency at coordinates The calculation is as follows: ； in, Image representing low-frequency coefficients Length and width and These represent the low-frequency coefficient images in and Low-frequency coefficient at the location; Visual significance of low-frequency coefficients Normalization is performed as follows: ； in, and These represent low-frequency coefficient images. The maximum and minimum values ​​among the visual saliency values; For low-frequency coefficient images respectively , Calculate the coordinates according to the formula above. Normalized visual significance at the location; Then, the fusion weights are calculated as follows: ； in, and These represent the low-frequency coefficient images respectively. , exist Normalized visual saliency at coordinates; and They represent , The corresponding fusion weight value; Finally, based on the visual saliency threshold VSM weighted fusion is performed to obtain the fused low-frequency coefficients. ; ； in, coordinates Low-frequency coefficients of the fusion process; Step Six: Apply the MSF maximization fusion rule based on threshold judgment to the high-frequency coefficients of the mean difference map and Gaussian difference map. We perform fusion, and let the resulting high-frequency coefficients be... ; A threshold-based MSF maximization fusion rule is used for high-frequency coefficients. To integrate, including: First, calculate the MSF of the high-frequency coefficients, as follows: ； in, Image representing high-frequency coefficients exist Spatial frequency values ​​at coordinates, where , These represent row and column frequencies, respectively. , For diagonal frequencies; for frequencies with high frequency coefficients Centered Window size , , , The calculation formulas are as follows: ； ； ； ； For high-frequency coefficient images respectively Calculate the coordinates according to the formula above. Spatial frequency value at that location; Then, based on the spatial frequency threshold Perform MSF scalar fusion to obtain the fused high-frequency coefficients. ; ； in, coordinates High-frequency coefficients of fusion , High-frequency coefficient images In coordinates High-frequency coefficients at the location; , These are the high-frequency coefficients. The corresponding MSF value; Step 7: Perform inverse NSST transform on the fused low-frequency and high-frequency coefficients to obtain the difference map after fusion; In this context, NSST represents unsampled shear wave, VSM represents visually salient mapping, and MSF represents improved spatial frequency.

Images