EXTRACTION OF EXTENDED SMALL-SCALE OBJECTS IN DIGITAL IMAGES

Detection and localization problem of extended small-scale objects with different shapes appears in radio observation systems which use SAR, infra-red, lidar and television camera. Intensive non-stationary background is the main difficulty for processing. Other challenge is low quality of images, blobs, blurred boundaries; in addition SAR images suffer from a serious intrinsic speckle noise. Statistics of background is not normal, it has evident skewness and heavy tails in probability density, so it is hard to identify it. The problem of extraction small-scale objects is solved here on the basis of directional filtering, adaptive thresholding and morthological analysis. New kind of masks is used which are open-ended at one side so it is possible to extract ends of line segments with unknown length. An advanced method of dynamical adaptive threshold setting is investigated which is based on isolated fragments extraction after thresholding. Hierarchy of isolated fragments on binary image is proposed for the analysis of segmentation results. It includes small-scale objects with different shape, size and orientation. The method uses extraction of isolated fragments in binary image and counting points in these fragments. Number of points in extracted fragments is normalized to the total number of points for given threshold and is used as effectiveness of extraction for these fragments. New method for adaptive threshold setting and control maximises effectiveness of extraction. It has optimality properties for objects extraction in normal noise field and shows effective results for real SAR images.


INTRODUCTION
The task of detection and localization of small extended objects in noisy images occur in the electronic surveillance systems using radars with SAR, infrared and laser systems, as well as television cameras (Gao, 2010;Gonzalez, Woods, Eddins, 2004).This task is relevant, because these facilities typically have an artificial origin and are of Prime interest.Upon detection, extraction and localization of such objects had substantial difficulties in building effective algorithms and structures processing, as in taken images there is intense and non-stationary background, also contains elements that are structurally similar to the signals, the signal/background is usually small, and the registered digital image has a low quality, small number of quantization levels, patency character and fuzzy borders of natural and artificial structures (rivers, roads, bridges, buildings).Statistics background is very different from a Gaussian, the distribution is clearly asymmetric, and the tails of the distributions like lognormal density normal or mixed (contaminated-normal), and when small volumes of samples are identified with difficulty.Such a character background virtually eliminates the use of the known methods of thresholding, since improper formation thresholds can cause loss of useful objects at a very early stage of processing.You cannot use traditional methods of detecting contours in images in order to highlight natural features (rivers, roads, borders, windbreaks, etc.), which are based on the formation of the spatial derivatives (gradients and Laplacians), because the result will be a significant growth impulse noise without visible effect for selection of quality circuit.
The basic principles that allow solving this difficult problem, are the location-based filtering, adaptive thresholding and selection of useful sites on the connectivity of neighboring pixels given the length of the useful structures (Gonzalez, Woods., Eddins, 2004).

PROBLEM STATEMENT AND METHOD OF OBJECT DETECTION
Suppose there is a classification problem, so that objects have to be extracted belong to certain classes.But these classes have no precise description and this is the object of investigation.Consider there is an image in digital form, containing useful objects, which have a relatively small length in relation to the size of the entire image and an arbitrary orientation.Shape of objects of interest can be linear or speckle, and their length is specified by specifying a maximum size or length of the object in pixels, and set the minimum and maximum bounds on the length of objects.The problem feature is that the emergences of small-scale objects of interest practically no effect on the integral characteristics of the image.
There are two examples of extended small-scaled objects in SAR images which are shown in Fig. 1 and Fig. 2.These objects are artificially highlighted with a white oval.First image contains two oriented linear objects and the second image contains two blob-like objects inside corresponding white ovals.
The general structure of digital image processing includes a prefilter, binary quantization (threshold processing), and subsequent morphological processing (Fig. 3).The input image after registration is submitted in digital form (two-dimensional array on a rectangular grid of points).
Figure 1.Small-scale oriented objects to be extracted The problem of automatic setting of the threshold in autonomous information and control systems is very important for segmentation (Akcay, Aksoy, 2007, Sezgin, Sankur., 2004, Weszka, Rosenfeld, 1978).(Gonzalez, Woods., Eddins, 2004, , Sezgin, Sankur., 2004).In our case, the threshold processing should depend on the results of binarization (Volkov, 2009a(Volkov, , 2009b)).The purpose of this paper is to study adaptive method threshold segmentation for the detection and selection of objects based on structural decomposition of a binary image into elementary, isolated objects, analysis of the impact of the threshold on the results of the decomposition, and algorithm development for installation and changes the threshold in accordance with the results of the decomposition.

PRELIMINARY FILTERING
Pre-filtering aims to improve the image and highlight the differences and boundaries.It is assumed that the useful objects always have a higher intensity relative to the background, otherwise it is necessary to invert the image.In this case we applied the differentiating filters (Laplacian type), which permit to use a global threshold for binary quantization exceeding the intensity threshold quantization.
When filtering oriented linear objects we used space-oriented mask filter of the following form (Fig. 4), which would have effectively allocate endpoints of the segments of unknown length.In this particular case the coefficients are . These masks should be used for searching fragments with different orientations for extraction all small-scale extended objects in the image.

THRESHOLD PROCESSING
This stage is very important.Incorrect threshold often results in irreversible losses of information.Suppose we are interested in objects with high intensity, and processing results in high level for pixels if threshold level is exceeded, and zero level otherwise.We can also see that for high thresholds bottom pictures has weak dependence on type of pre-filtering processing.The main idea is to set threshold level according to segmentation results.For this purpose hierarchy of isolated fragments is proposed, and effectiveness of extraction is inserted as indicator of segmentation degree.It may be used for threshold setting and control.

HIERARCHY OF ISOLATED FRAGMENTS IN BINARY IMAGE
Our aim is to find out attributes of image which characterize extension properties of objects and allow us to control threshold level for qualitative segmentation.Suppose we have binary image after threshold processing.The concept of characterizing mask for isolated fragments is introduced for analysis of segmentation results.It relates to definition of continuity and adjacency of pixels, here usual definitions are used (Gonzalez, Woods., Eddins, 2004, Volkov, 2009a, Volkov 2009b).
Fragments are isolated if they have no mutual pixels.Suppose isolated fragment consists of several adjacent pixels on the binary image.Hierarchy of fragments is their ordering and may be obtained by the choice of characterizing masks to be considered.Masks 2x3, 3x5, 3x7, 5x7 and other similar masks characterize horizontal fragments; masks 3x2, 5x3, 7x3, 7x5 describe vertical fragments.
The choice of set of characterizing mask allows us to obtain different attributes for describing fragments with different extensions and orientations.
The simplest hierarchy uses only square characterizing masks.Then we have 1x1, 2x2, 3x3, 5x5, 7x7 and so on, as characterizing masks of extensive fragments with increasing extensions.This hierarchy does not take into account orientations of fragments.

THRESHOLD SETTING AND CONTROL BY THE USE OF CONNECTED FRAGMENTS IN NOISE IMAGE
It is easy to verify that the number of extracted small-scale objects of the given type in binary image is small at both very low and high thresholds.Thus, with some intermediate value of the threshold number of such objects is maximal.(2) and probability for any fragment with characterizing mask 2x2 is and so on were calculated and represented on the top of Fig. 14.Numbers of lines correspond to characterizing masks 1x1, 2x2, 3x3, etc. Extraction efficiency for isolated points (line 1), fragments with characterizing mask 2x2 (line 2), and mask 3x3 (line 3).Lines for connected fragments have evident maxima at the threshold value 1.3.This is the best threshold for extracting small connected fragments from Gaussian noise field.These results are confirmed by simulation curves represented on the bottom of Fig. 14.

EXTRACTION OF SMALL-SCALE OBJECTS FROM REAL SAR IMAGES
The results of the extraction and allocation of small-scale objects on real radar images presented in Fig. 16

CONCLUSIONS
New method of extraction and allocation small-scale objects is described which is based on pre-filtering, adaptive thresholding and morphological selection.It allows extracting small-scale objects with different extension and orientation.Adaptive method for threshold setting and control has optimal property which was checked by modeling Gaussian field with extensive object region.Thresholds obtained are settled near the value of optimal maximal likelihood threshold for detection of shift on Gaussian field.

Figure 2 .
Figure 2. Small-scale blob-like objects to be extracted

Figure 6 .
Figure 6.The output of pre-filtering of image in Fig. 1

Figure 7 .
Figure 7.The output of pre-filtering of image in Fig. 2

Figure 8 .
Figure 8. Binary image after low threshold level

Figure 9 .
Figure 9. Binary image after intermediate threshold level

Figure 11 .
Figure 11.Gaussian noise field (top picture) and the result of binarization (bottom picture) with threshold level equals to 1.5

Figure 12 .
Figure 12.Isolated points extracted (top picture) and remaining small-scale objects (bottom picture) after thresholding of Gaussian noise field

Figure 13 .
Figure 13.Hierarchy of small isolated fragments

Figure 14 .
Figure 14.Extraction efficiency for isolated points (line 1), fragments with characterizing mask 2x2 (line 2), and mask 3x3 (line 3) and Fig 17 for the two images from Fig. 1 and Fig. 2 accordingly.Top pictures represent outputs of binary quantizers which threshold were set for maximum extraction efficiency for small connected fragments.Objects of interest are highlighted by the ovals.Dependency plots of extraction efficiency upon threshold values are presented in the middle row of pictures.Numbers 1, 2 and 3 correspond to connectivity of pixels in fragments.Pictures on the bottom show objects of interest extracted by the use of length selection.

Figure 15 .
Figure 15.Processing structure with adaptive threshold setting and control

Figure 16 .
Figure 16.Extraction of small-scale objects from the image in Fig. 1

Figure 17 .
Figure 17.Extraction of small-scale objects from the image in Fig. 2 Setting of threshold is the main problem for qualitative segmentation.Low threshold level gives much noise, and the following processing becomes inefficient, too high level results in destroying of useful objects which may be split up to small fragments.The best threshold should be set after analysis of segmented fragments with the use of some quality indicator for extraction and segmentation.It is desirable to get some attributes which characterize the quality of segmentation.The simplest attribute is proposed here as the number of points at each step of extraction.It should be normalized to the initial points in binary image and represents the effectiveness of extraction at corresponding step.It is worth noting that threshold setting is dependent on prefiltering processing.The general idea of threshold setting and control via results of segmentation is admissible for different pre-filtering algorithms.It also may be used for local thresholds in sliding windows.