FIELD OF INVENTION
The present invention is related to one of the type of SAR i.e. spotlight SAR 5 where multiple SAR images are captured of the exact same region/ spot of the earth surface within seconds of the time interval. The captured SAR images are inherently speckled in nature. This scenario generates multiple similar SAR images. Since the SAR images are inherently speckled, so the noise variance of those multiple SAR images are different. More principally, the invention is related 10 to removal of speckle noise present in those multiple similar spotlight SAR images. The speckle noise is mainly introduced in the image acquisition phase. Radar imagery is especially used during monsoon season. Since the spotlight SAR images are captured within seconds of the time interval, therefore the chances of disturbance of SAR sensor or antenna is high which can generate noise in the 15 captured SAR images. Another factor responsible for speckle introduction in the
3
spotlight SAR image is the consistent interaction of the transmitted high-frequency radar signal with the target region. This consistent interaction creates random constructive and destructive interference which results into multiplicative kind of scattering phenomenon known as speckle noise. The available image restoration models are additive in nature. So, the multiplicative nature conversion 5 is desirable to additive through which the existing image restoration models can be used efficiently. There is multiple spotlight SAR applications like agricultural land monitoring, wave forecasting, marine climatology, regional ice monitoring, ship detection in the coastal regions and field of tropical forest monitoring.
BACKGROUND & PRIOR ART 10
A solution to an serious issue in the spotlight SAR type is raised, where multiple images are captured of the exact region (spot) from different locations at a difference of seconds of time. The SAR images captured are inherently speckled in nature. Here the moving SAR system like aircraft etc. continuously steers at the exact scene (generally on the earth) such that same target region is illuminated and 15 captured from different locations within seconds of time. This mapping captures multiple SAR images which are individually different as per nature but of the same target spot (region). The spotlight SAR provides fast time complexity and enhanced azimuth resolution than stripmap SAR but the captured image area is smaller. In this mapping when aircrafts move forward, it captures the image by 20 transmitting the high-frequency radar waves from the antenna attached to its SAR. These high-frequency radar waves hit the target on the ground and have bounced off objects on the ground. The antenna has the ability to collect the backscatter information came from the ground. These backscatter pulses comprise of information including the total distance toured by the pulses to make the round 25 trip to and from the plane and it also contains the information about the movement of the SAR if it is moving in the direction of or against from the object on the ground. It also contains the information about travel time of each radar wave and also evaluates the time interval between the adjacent radar waves. All the information is returned back to the antenna in the form of data, which is later 30
4
processed by the spotlight SAR. Since the information is of region of small area, therefore the computational cost and processing time is smaller than other SAR types. When high-frequency radar waves hit the target ground, this interaction is consistent which initiates random constructive and destructive nosiness resulting into multiplicative kind of noise known as speckle noise all over the image. 5 Therefore it is said that the captured SAR image is inherently speckled. Similarly, this process is repeated by the aircraft from another location and it captures the another image of the same region, which is also speckled inherently. This automatic inclusion of speckle noise in the SAR images causes detection and segmentation challenges in later satellite image processing steps. Therefore there 10 is a need of despeckling method based on multiple captured SAR image information, so that maximum enhancement can be achieved and it exhibits a single clean and despeckled SAR image. The despeckling method based on fusion is used and experimentally tested on the simulated SAR images. Since the captured SAR images are two different images of the same scene, therefore their 15 noise variances are also different. The experimentation is executed by introducing speckle noise in clean SAR images for performance evaluation. For easy and experimental understanding, two original SAR images of the same location are taken. The noise is introduced in them and is assumed that the noise introduced during the image acquisition process. 20
There are many despeckling techniques and filters available in spatial as well as frequency domain to handle speckled SAR images by preserving the major information in the image like edge, texture, homogeneous and non-homogeneous regions of the image. Some standard and conventional method for despeckling purpose in the field of satellite imagery are, frost filter, Kuan filter, Kuwahara 25 filter, lee filter, mean filter and median filter. Other developed schemes under homomorphic and non-homomorphic filtering are effective and adaptive. It is observed that Bayesian approaches in transform domain show better results than Bayesian approaches in the spatial domain. But there some non-bayesian approaches that give as better results as Bayesian approaches in the transform 30 domain. Order Statistics and Morphological Filters are helpful in preserving
5
edges, retaining texture and smoothing the noisy background, but thicker objects left unprocessed. This method is not specifically built for speckle noise model but still provides fair results in some cases. Anisotropic diffusion is a popular technique that retains significant parts mainly edges, lines, and other fine details. Speckle reduction using anisotropic diffusion exploits sudden occurring 5 coefficient of variation. It shows better results than conventional methods in the terms of variance minimization, mean preservation, and edge localization. In despeckling based on compressed sensing it is a known that in order to obtain a good quality image, multiple degraded images can be merged. Keeping this in mind, compressed sensing is employed to get multiple SAR images from a single 10 SAR image. Other non-bayesian approaches like bilateral filtering, sigma filter and non-local filtering which also gives satisfactory results in terms of visual appearance and edge preservation. Homomorphic and non-homomorphic filtering are the two Bayesian methods in transform domain that provides best schemes in the wavelet domain. Homomorphic filtering is used to remove the multiplicative 15 noise using log and exponential operations, while non-homomorphic filtering is less frequently seen in the literature due to its complexity of directly handling the multiplicative noise. Homomorphic filtering is in much use since last two decades as after transforming multiplicative noise to additive, other additive noise models can be used easily and effectively to handle the situation, it is an easy method to 20 understand, while non-homomorphic filtering methods directly work upon multiplicative noise. This method is comparatively difficult to work upon and is less effective too. Basically, homomorphic filtering is used for improving non-homogeneous illumination in images. Classical hard and soft thresholding methods were implemented in. The undecimated wavelet transform and the MAP 25 standard have been implemented in the issue of SAR image despeckling.
OBJECTIVE OF THE INVENTION
The principal objective of the present invention are:
 To perform despeckling of speckled spotlight SAR images using fusion.
6
 Removal of granular structured speckle noise from multiple speckled spotlight SAR images using fusion concept.
 To resolve the problematic issue in spotlight SAR imagery.
 Preservation of edges, texture, homogeneous and non-homogeneous regions of agriculture land. 5
 To enhance the quality of the despeckled image by using the similarity (CC) and dissimilarity (MMSE) features in the despeckling process.
BRIEF DESCRIPTION OF DRAWINGS
The present invention eliminates the granular structured speckle noise (1,12) from 10 multiple SAR images (1, 12) which occurred during image acquisition and fuses them into a single enhanced and high quality despeckled SAR image (11). Before performing any type of satellite image processing operation like segmentation, morphology, recognition etc, SAR image enhancement and despeckling is the first preprocessing step to be performed. The present invention uses homomorphic 15 filtering (2, 10) in order to handle multiplicative nature of speckle noise (1, 12). It takes two speckled SAR images (1, 12) as input and applies homomorphic filtering operation (2, 10) on it. Later 2D-DWT (3, 13) is applied on both of them in parallel. The approximate parts (4, 14) are not processed. The detailed part (5, 15) are thresholded using Bayesian (6) and Bivariate thresholding (16) methods. 20 The approximate parts (4, 14) are fused on the basis of correlation (7, 8). The high-frequency detailed parts (5, 15) are fused using minimum mean square error (MMSE) and correlation coefficient (CC) (17, 18, 19). This process yields an enhanced low (8) and high-frequency wavelet coefficients (19) which are later directed to IDWT (9) and then to exponential transform (10). This whole 25 procedure finally provides the despeckled output image (11).
DETAIL DESCRIPTION
The invented framework (Fig 2) is designed for despeckling the spotlight SAR images (1,12) using fusion (8, 19). Fig 1 describes the problematic issue raised in the spotlight SAR type. In brief, the issue raised is that the aircraft (102) from 30
7
location 1 (101) transmits the high-frequency radar signals (109) on area A (107) to capture it. These high-frequency radar signals (109) hit the ground (107). After hitting the target (107), a portion of the energy signals (110) is bounced back to the antenna of the aircraft(102). This energy signal (110) contains various information about target object (107) and aircraft (102). Using this information 5 (110), after complex computation, a SAR image 2 (108) is created. This coherent interaction of high-frequency radar signals (109, 111) with the target area (107) causes constructive and destructive interference resulting to the introduction of speckle noise in the captured SAR images (108, 106). The same process is repeated by the aircraft (105) from the location 2 (104) and SAR image 1 (106) is 10 created there. Since it is known that the due to coherent interaction, the SAR images (108, 106) are inherently speckled in nature. Since each captured SAR images (108, 106) is speckled up to some noise variances, so some or more detailed information is missing in every image. The invented framework (Fig 2) resolves this issue by working on two different SAR image (1, 12) and later 15 enhances it using fusion concept (8, 19).
The invented framework (Fig 2) is designed and experimented on the two different SAR images (1, 12) of the same scene (107) but the framework is adaptive enough to work on multiple SAR images. The despeckling framework works as follows: It takes two different SAR images A (1) and B (12) of the same 20 scene (107) and let their noise variance (σ) be m (1) and n (12) respectively. The framework is based on homomorphic filtering (2, 10). The log transform (2) is applied to both speckled images A (1) and B (12) in parallel. Then db8 based 2D-DWT (3, 13) is applied on both log transformed speckled SAR images in parallel.
25
Sign: -
Name: - SHREE, RAJ
Address: - B-1/1-4, DEPARTMENT OF INFORMATION TECHNOLOGY, BABASAHEB BHIMRAO AMBEDKAR UNIVERSITY, VIDYA VIHAR, LUCKNOW (226025), UP, INDIA
Sign: -
Name: - SINGH, PRABHISHEK
Address: -DEPARTMENT OF INFORMATION TECHNOLOGY, BABASAHEB BHIMRAO AMBEDKAR UNIVERSITY, VIDYA VIHAR, LUCKNOW (226025), UP, INDIA
Sign: -
Name: - PANDEY, RAVI PRAKASH
Address: - 5/5/21A, NEAR UDAYA PUBLIC SCHOOL, ITI CHAURAHA, PANCHKOSHI PARIKRAMA MARG, FAIZABAD, UP, INDIA-224001
Sign: -
Name: - SHUKLA, VIVEK
Address: -DEPARTMENT OF INFORMATION TECHNOLOGY, BABASAHEB BHIMRAO AMBEDKAR UNIVERSITY, VIDYA VIHAR, LUCKNOW (226025), UP, INDIA
Sign: -
Name: - DIWAKAR, MANOJ
Address: - DEPARTMENT OF CS & IT, UTTARANCHAL UNIVERSITY, DEHRADUN, 248007, INDIA
8
The SAR image A (1) is decomposed into the approximate part (LL’) (4) and detailed part (LHm, HLm, HHm) (5). Similarly, the SAR image B (12) is also decomposed into the approximate part (LL’’) (14) and detailed part (LHn, HLn, HHn) (15). The framework uses wavelet thresholding method (6, 16). The approximate parts (LL’ and LL’’) (4, 14) are not processed and remained 5 untouched for a while. The first detailed part (LHm, HLm, HHm) (5) is thresholded using bayesian thresholding method (6), while the second detailed part (LHn, HLn, HHn) (15) is thresholded using bivariate thresholding method (16). The speckled detailed parts (LHm, HLm, HHm) (5) and (LHn, HLn, HHn) (15) are now thresholded and despeckled and denoted as (HL’, LH’, HH’) (6) and 10 (HL’’, LH’’, HH’’) (16) respectively.
The concept of fusion (9, 19) is introduced in the SAR image despeckling to enhance the quality of the output image. The low-frequency wavelet coefficients (LL’ and LL’’) (4, 14) are fused based on correlation (7, 8). The LL’ (4) and LL’’ (14) are compared and correlated pixel-to-pixel using CC (7) and based on this the 15 fusion operation is performed (8). Firstly, the threshold value is calculated using this approach: Non-overlapping block-wise CC’s (7) are calculated from these low-frequency coefficients (LL’ and LL’’) (4, 14) using 3 × 3 mask size. The average of all the CC (7) is evaluated and is decided as the threshold value. In the local correlation based fusion strategy (8) of low-frequency coefficients (LL’ and 20 LL’’) (4, 14), a non-overlapping block-wise CC (7) are calculated from these low frequencies (4, 14) using 3 × 3 mask size. Here, the acquired CC (7) is compared with the evaluated threshold value. If the CC (7) is less than or equal to the threshold value, then the maximum operation is executed.
25
Sign: -
Name: - SHREE, RAJ
Address: - B-1/1-4, DEPARTMENT OF INFORMATION TECHNOLOGY, BABASAHEB BHIMRAO AMBEDKAR UNIVERSITY, VIDYA VIHAR, LUCKNOW (226025), UP, INDIA
Sign: -
Name: - SINGH, PRABHISHEK
Address: -DEPARTMENT OF INFORMATION TECHNOLOGY, BABASAHEB BHIMRAO AMBEDKAR UNIVERSITY, VIDYA VIHAR, LUCKNOW (226025), UP, INDIA
Sign: -
Name: - PANDEY, RAVI PRAKASH
Address: - 5/5/21A, NEAR UDAYA PUBLIC SCHOOL, ITI CHAURAHA, PANCHKOSHI PARIKRAMA MARG, FAIZABAD, UP, INDIA-224001
Sign: -
Name: - SHUKLA, VIVEK
Address: -DEPARTMENT OF INFORMATION TECHNOLOGY, BABASAHEB BHIMRAO AMBEDKAR UNIVERSITY, VIDYA VIHAR, LUCKNOW (226025), UP, INDIA
Sign: -
Name: - DIWAKAR, MANOJ
Address: - DEPARTMENT OF CS & IT, UTTARANCHAL UNIVERSITY, DEHRADUN, 248007, INDIA
9
In maximum operation, fusion is implemented by choosing the largest values from both of the transformed coefficients. Or else, the average operation is applied, which calculates the average value using both of the low-frequency coefficients (4, 14) to perform fusion (8). The resultant enhanced low-frequency coefficients are LLnew (8). 5
The next fusion is performed with high-frequency coefficients (HL’, LH’, HH’) (6) and (HL’’, LH’’, HH’’) (16). The method of calculating the threshold value is same based on CC (18). Here pixel-to-pixel MMSE (17) as well as CC (18) is evaluated over high frequency coefficients (HL’, LH’, HH’) (6) and (HL’’, LH’’, HH’’) (16) which evaluates and analyzes the degree of similarity (CC) (18) and 10 dissimilarity (MMSE) (17) over (HL’, LH’, HH’) (6) and (HL’’, LH’’, HH’’) (16). Based on CC (18), MMSE (17) and the threshold value, fusion (19) is performed. The CC (18) shows the similarity factor and MMSE (17) shows the dissimilarity factor. Since the range of CC (7, 18) is [0-1], therefore to utilize the MMSE value in fusion operation, the original MMSE value (17) is normalized in 15 [0-1] range. After this, the average of CC (18) and (1-MMSE) (17) is evaluated and named it as a correlated mean square error (CMSE) (19) and then the obtained CMSE (19) is compared with the evaluated threshold value. If the CMSE (19) is less than or equal to the threshold value, then the maximum operation is executed. In maximum operation, fusion is implemented by choosing the largest values from 20 both of the transformed coefficients. Or else, the average operation is applied, which calculates the average value using both of the high-frequency coefficients to perform fusion (19).
25
Sign: -
Name: - SHREE, RAJ
Address: - B-1/1-4, DEPARTMENT OF INFORMATION TECHNOLOGY, BABASAHEB BHIMRAO AMBEDKAR UNIVERSITY, VIDYA VIHAR, LUCKNOW (226025), UP, INDIA
Sign: -
Name: - SINGH, PRABHISHEK
Address: -DEPARTMENT OF INFORMATION TECHNOLOGY, BABASAHEB BHIMRAO AMBEDKAR UNIVERSITY, VIDYA VIHAR, LUCKNOW (226025), UP, INDIA
Sign: -
Name: - PANDEY, RAVI PRAKASH
Address: - 5/5/21A, NEAR UDAYA PUBLIC SCHOOL, ITI CHAURAHA, PANCHKOSHI PARIKRAMA MARG, FAIZABAD, UP, INDIA-224001
Sign: -
Name: - SHUKLA, VIVEK
Address: -DEPARTMENT OF INFORMATION TECHNOLOGY, BABASAHEB BHIMRAO AMBEDKAR UNIVERSITY, VIDYA VIHAR, LUCKNOW (226025), UP, INDIA
Sign: -
Name: - DIWAKAR, MANOJ
Address: - DEPARTMENT OF CS & IT, UTTARANCHAL UNIVERSITY, DEHRADUN, 248007, INDIA
10
The resultant enhanced high-frequency coefficients are (LHnew, HLnew, HHnew) (19). Now the enhanced low (LLnew) (8) and high-frequency wavelet coefficients (LHnew, HLnew, HHnew) (19) are obtained. The inverse DWT (IDWT) (9) is performed using (LLnew) (8) and (LHnew, HLnew, HHnew) (19). Now the anti-log operation i.e. ‘exponential operation’ (10) is applied to the output of IDWT 5 (9). This yields the final clean despeckled SAR image (11).

The despeckling SAR image method comprising:
1. It provides a homomorphic and unique framework for obtaining the best despeckling results in spotlight SAR imagery using fusion. 5
2. The framework has the capability to obtain the high quality and enhanced image from multiple degraded images. It provides an adaptive method which can be used in any denoising application for the similar type of problem.
3. The framework is computationally efficient, easy to understand and 10 implement with best results.
4. The solution provided by the framework to the problem of spotlight SAR imagery is unique and never done before.
5. The framework used hybrid thresholding (Bayes and Bivariate) methods on high-frequency wavelet coefficients. 15
6. The framework fuses the low and high wavelet coefficients of two different thresholded methods using CC and MMSE based on a newly created mathematical formula.
7. The mathematical formula for fusion is based on similarity and dissimilarity factor. It uses both the factors for optimal results by 20 introducing the similarity features and removing the dissimilarity features during the fusion procedure.