ارائه فيلتر كاهش نويزاسپكل روي تصاوير رادار با روزنه تركيبي

Proposal speckle reduction algorithm for SAR images

وجود اسپكل در تصاویر رادار امری اجتناب ناپذیر است. نویز اسپكل یك آشفتگی نقطه‌ای است كه معمولا به عنوان نویز ضرب‌شونده در تصاویر تك‌پلاریزه مدل می‌شود. این نویز كه وابسته به سیگنال است به دلیل نوسانات فاز سیگنال‌های بازگشتی امواج الكترومغناطیسی بوجود می‌آید كه به صورت نقطه نقطه ظاهر می‌شود. حضور اسپكل، تفسیر و آنالیز تصویر را پیچیده‌تر می‌كند و باعث كاهش دسترسی به اطلاعات تصویر می‌شود لذا انتخاب الگوریتم مناسب كاهش نویز اسپكل حائز اهمیت است. در مقاله حاضر ابتدا به توضیح فیلترهای متداول كاهش اسپكل مانند: میانگین، كوان، فراست، لی، لی بهبودیافته و گاما-مپ پرداخته شده، سپس بر طبق ویژگی‌های آماری اسپكل و ویژگی‌های بافتی تصویر SAR یك الگوریتم تطبیقی برای كاهش نویز اسپكل با پنجره تغییرپذیر از لحاظ بعد بر پایه انحراف معیار نسبی پیشنهاد شده است. برای بررسی و ارزیابی الگوریتم پیشنهادی و سایر فیلترهای متداول، از یك تصویر شبیه‌سازی شده چهار-دید SAR و دو تصویر واقعی، اخذ شده از دو سنجده هوابرد مختلف، استفاده شد. نتایج تجربی نشان‌دهده عملكرد رضایت‌بخش الگوریتم پیشنهادی در حذف اسپكل، همراه با حفظ بسیار خوب ویژگی‌های لبه، اهداف نقطه‌ای و میانگین تصویر نسبت به سایر فیلترها می‌باشد. به عنوان مثال شاخص‌های ENL و SNR برای برای ناحیه 1 از تصویر اول به ترتیب 17.22 و 7.91 بود كه بالاتر از مقدار سایر فیلترهای متداول بود.

The existence of speckle in Radar images is an inevitable occurrence. The speckle noise is a granular disturbance that models often as a multiplicative noise in single-channel SAR images. This noise that dependent on signal, introduces for variations of phase of returned signals that appears as point-point pattern. Being speckle, make complex more the explanation and analysis of images and also decrease the access to the image data, so it is important which appropriate speckle-reduction algorithm should be choosing. therfore, a section is allocated to the introduction of speckle noise model. It provides some important facts about how the speckle formed and explains which probably density function is followed by the amplitude and the intensity image. Several adaptive filtering methods have been discussed to deal with issue in this paper such as: Mean, Kuan, Frost, Lee, enhanced Lee and Gamma-MAP. Then according to statistic characteristics of speckles and texture characteristics of SAR images an adaptive speckle-reduction algorithm with size-changeable window based on relative standard deviation have been put forward. This new algorithm uses a moving window like other typical filters, but its window is divided to four smaller windows that every of them is called subwindow. The relative standard deviation of every subwindow is used as compare factor. If all subwindows are in homogenous region the mean filter is applied on whole initial window. And if any subwindow has the edge or high-frequency information, it must be omitted from proceeding process. The proposed filter is size-changeable because, if all subwindows are not accepted by the rules, the filter need to reduce the size of initial window to provide a region for filtering process. This paper benefits some worthy indices to demonstrate the ability of proposed filter against common filters such as equivalent number of looks (ENL), speckle suppression and mean preservation index (SMPI), edge save index (ESI), mean square error (MSE) and signal to noise ratio (SNR). To analyze the proposal algorithm and other common filters, used a simulated four-look SAR image and two real SAR images i.e. Flevoland dataset from AirSAR airborne SAR sensor and Oberpfaffenhofen dataset from ESAR airborne SAR sensor, respectively. The simulated SAR image is determined according to multiplicative model and gamma distribution. It is used to show a primary evaluation of proposed filter and other filters. In the first step, the results are satisfying. For example, the indices of standard deviation (SD) and ELN for the proposed filter are 39.53 and 192.59 that in comparison to other filters are gratifying and agreeable. In the next step, the all filters that are described in this paper are applied on Flevoland dataset. As the experience results show, the proposal algorithm has a satisfying performance in removing speckle noise along with very good saving edge characteristics, targets scene and mean of image toward usual speckle-reduction filters. For instance, the indices of ENL and SNR for area 1 of first image are 17.22 and 7.91, respectively, that are highest values between other common filters. To survey more precisely, the mean and the enhanced Lee filters are selected to compare to the proposed filter by second real dataset in the ways of ability to remove speckle, preserve edge and maintain point target.