Modeling and Information Processing
ATR - Automatic target recognition
Detecting objects in complex images
The problem: detecting targets (manufactured moving objects) in satellite images coming from high resolution sensors in the visible band.
Introduction
The size of satellite images (from several hundred Mega-pixels today, to several Giga-pixels in the near future) bars the direct use of sophisticated analysis methods on any image. Therefore, an effective exploration strategy must be set up to control the computing cost.
To do this, we have uncoupled Detection and Recognition into two phases:
- Focus / Detection Phase: This is a question of focusing very quickly on the "salient" areas of the image (i.e. the areas that could contain targets) by extracting simple characteristics. Each pixel is attributed a class: 1 for "salient", 0 if not. Thus, the result is a binary segmentation of the image. The algorithms employed must be low cost or easy to execute in parallel (such as filter banks). In this phase, a probability of detection Pd = 1 must be obtained. The false alarm rate authorized can be relatively high.
- Detection/Recognition: The aim of this phase is to eliminate the maximum of false alarms (reduce the false alarm rate whilst keeping Pd high). The processing planned can be much more refined because it is restricted to small areas.
The work presented here concerns the Focus phase.
Focus phase
The procedure must provide answers to numerous problems: variability of shapes and contrasts, invariance to the orientation of targets and variability of the size of targets. Because we have limited the complexity of the processing, we use a "minimal" representation of targets. A target is a set of contour elements in a limited spatial area with preferential orientations, and contrasted in relation to the background. Two local parameters seem important for detection: contrast and scale. Scale is a hidden parameter that constrains the spatial organization of the local contrasts.
We have created a salient region detector that is capable of managing objects of varied sizes. This detector is based on a probabilistic model that combines two types of multi-scale characteristics calculated locally: a series of characteristics sensitive to contrast and a series of characteristics that give information about the local scale.
Contrast detectors and local estimation of the scale
The local contrast descriptors must include a scale parameter and be sensitive to spatially extensive structures. To do this we have used a simple technique which consists of calculating, for each site s of the image, the responses of a local detector over a spatial neighborhood V e S with a size e. This size e is our descriptor's size parameter. Two contrast detectors based on this principle have been used. They are built from oriented-edge detectors (Gabor filters for example). In practice, these detectors work well if their scale parameter e is adapted to the size of the structure to be detected.
In order to automatically detect objects of variable sizes, this scale parameter must be estimated locally for every site s. Lindeberg (1998) give a definition of the scale adapted to site s: it is a local extremum (on the scales e) of a multi-scale characteristic calculated at site s in the image. This adapted scale is then used to set the parameters of a contrast detector. We have used the local entropy (Kadir, 2001) as a scale selection indicator characteristic.
Combining the scale and contrast information: Probabilistic model
We wish to attribute a class As (1 or 0, salient or not) to each pixel s. To do this, we consider As and the scale s to be random variables. Each pixel's class is attributed by comparing the conditional probability P(As = a|Ys) to a detection threshold T, where Ys = (Des, Hs) is the set of contrast and scale data calculated at a pixel.
As the scale is considered to be a hidden variable, we have defined a combination scheme in the form:
The probability laws for As and Des are described by symbolic logic models. This combination model is characterized by a small number of parameters that can be learnt using an "Expectation-Maximization" type algorithm.
Results
The method was tested on a problem of aircraft detection in an aerial airport image (ortho-image accessible at ortho.mit.edu/nsdi/draw-ortho.cgi?image=241898). The experiments gave good results: a high rate of detection can be achieved even when using very few learning examples. However, the salient areas detection results must be completed by a false alarm separating stage.
Figure 1 - Example of an image in which to detect objects of interest (aircraft). The circles indicate the centers and sizes of the salient areas detected.
Figure 2 - Map of probability of the presence of salient objects.
Figure 3 - Salient areas obtained after thresholding and morphological opening of the salient probability map
References
[1] T. Kadir, M. Brady, Scale Saliency : A Novel Approach To Salient Feature And Scale Selection, IJCV 2001
[2] T. Lindeberg, Feature Detection with Automatic Scale Selection, IJCV 1998
[3] B. Chalmond, B. Francesconi and S. Herbin. Using Hidden Scale for Salient Object Detection. Prépublication du Centre de mathématiques et de Leurs Applications, ENS de Cachan, Nº 2004-18.
www.cmla.ens-cachan.fr/Cmla/Publications/2004/CMLA2004-18.pdf
