The basic advantage of air reconnaissance of mine fields is their high efficiency and rather low cost. Among drawbacks of this method is impossibility to detect mine fields consisting of 1-2 mines since they can be confused easily with other objects located on the ground surface. Another limitation emerges when it is necessary to detect mines installed in the ground. In this case the possibility of detection depends on how much a buried mine violates the ground surface.

In solving the problem of air reconnaissance of mine fields, an algorithm consisting of following stages is used:

At the preliminary stage salient points on a given image are detected. A part of them should belong to the objects of interest. At this stage, independent of scale fast algorithms are applied which based on the convolution and consider both values of brightness of points on the image and its derivatives of various orders. At the given stage prior information about the interesting objects can be used to make choice of the applied filter.

The encoding is viewed at as attributing to each of the selected points a multidimensional vector, every component of which is the result of applying a custom filter. Such a vector is also referred to as a local descriptor. Prior available information is used in customizing a filter. A local descriptor should capture the most informative attributes which are used for identification of interesting objects. Among these attributes can be contours, textural heterogeneities, distribution of intensity and its spatial derivatives. The choice of a local descriptor, thus, is specific to a problem at hand. Some basic requirements driving this choice are invariance to a class of given transformations and variable illumination conditions, stability at the demanded noise level.

The stage of classification consists of two steps. At the first step the total set of vectors and corresponding points is divided into two classes: belonging to a minelike object or belonging a background. The decision is made by comparing the values of components of a local descriptor that are responsible for a certain type of symmetry, peculiar to the majority of mines, with a predefined threshold. The second step of classification is to divide the set of minelike objects into distinguished subclasses corresponding to different types of mines.

An example of image processing results obtained by exploiting contours and symmetry of the objects

At the final stage of processing the position of found mines is analyzed and decision is made whether the distribution of mines belongs to a type of known minefields. Known minefields are characterized by statistical distributions of mines relevant to a way of installing a minefield: from air, by volley fire, and ground minelayers. To define the extent of the minefield the set of found mines is contoured and exact geographical coordinates are defined.

The mine field represented in the above picture was installed in the ground at a depth about 10 cm a few months before acquiring the image. Unmasking attributes of this minefield were the visible changes in vegetation cover. In addition to images taken in visible spectrum, infra-red or radio-band images can also be used. The described methodology can also be applied to images in these bands with little modification.