In this project we are proposing a novel approach to the characterization of sea surface. This approach enables us to identify the sea surface in a given image.
Abstract
In this project we are proposing a novel approach to the characterization of sea surface. This approach enables us to identify the sea surface in a given image, and by doing so we can establish the existence of objects on it.
Background
When analyzing a video sequence, in order to allow an algorithm to accumulate information, a link must be established between consecutive images. Normally, the link is established by assuming the surroundings to be static. However, in the case of the sea surface, this assumption is altogether unreliable: the movement of the waves and the lack of fixed objects on the sea surface prevent us from referring to the sea as of having a static nature. Therefore, trying to detect suspicious objects solely on the basis of movement is bound to fail. A need for the characterization of the sea surface arises, a characterization that would give us the ability to identify the sea, and hence, objects on it that do not share the same characteristics.
Basic Approach
- The problem was divided into three parts:
- finding the boundaries of the sea in the picture, namely the horizon or the shoreline
- characterizing the sea surface, and thus identifying non standard objects on it
- segregation of foam and wake from actual targets
The problem of finding the horizon \ shoreline was tackled by the use of a variant of the Hough transform dedicated to finding straight lines in an image. The procedure considers each pixel’s direction and distance from the center of the image in order to establish which line has the most pixels in it. We assume here that the shoreline is the most dominant straight line in the picture.
Shoreline Algorithm example: the blue line on the original image represents the result.
After having established the boundaries of the sea, we are now free to characterize it. We chose to regard the sea as of having a certain texture, and therefore tried to characterize it using texture descriptors.
The basic idea involves learning the characteristics of a small peripheral area assumed to be the sea surface, and trying to find the same features throughout the whole area as defined by the shoreline algorithm.
Sea Learning Algorithm example : red area is identified as the sea surface. Note that the location of the shoreline was determined earlier.
After having identified all non typical sea surface objects we are now facing the task of segregating the target from its wake or from an occasional foam. This was achieved using morphological operations, under the assumption that wake and foam have a somewhat less complex shape than the target, and a smaller ability to retain the same shape over time.
Segregation Algorithm example: on the left is the original image. On the right we can see the same image, a few frames ahead. Little red crosses mark the most correlated complex areas. Note that a relatively complex wake does not retain its shape over time.
Tools
All project code was written in Matlab.
Conclusions
The algorithm provides automatic target recognition capability at sea. It makes very few confining assumptions about the target size, colour and shape, and is therefore not confined to a certain type of target (in certain cases this can prove to be a handicap, when the image contains different non-target objects, such as a small island).
Furthermore, it should be noted that the algorithm is based on tracking anomalies in the sea surface texture, and therefore can detect targets that are within the texture boundary. It does not deal with the case in which the target is above the horizon.
Acknowledgments
First and foremost, I would like to thank my instructor Guy Caspary, for his helpful advice and guidance. I would also like to thank the Ollendorff Minerva Center Fund for its continuous support that facilitated this research.