During the three-year period 1997 to 2000, a computer software system was developed to assist biologists in identifying individual dolphins in field photographs. The system maintains a database of dorsal fin images, and matches new images to the individuals in the collection. The system is in use in the laboratories of colleagues. During the three year period of this award, new data acquisition, feature extraction and indexing techniques will be introduced, improving the system's performance and extending its use to additional biological species.
Many field photographs are unusable because of poor photographic quality, caused by motion blur, splashes, or glints of reflected sunlight that obscure part of the fin's boundary. This problem will be addressed by using digital video data acquisition. Sequential frames showing the same fin will be combined, reducing noise and interference and increasing the likelihood of obtaining usable images.
The analysis of the shape of the trailing edge of the dorsal fin will be extended by using spatially coarse descriptors that capture the overall distribution of notches. These descriptors will be computed automatically and will capture properties such as "has deep notches" and "most notches are near the tip." These properties will be more robust than matching the exact edge shape, and will rapidly narrow the search and reduce errors.
A variant on the kd-tree search strategy, based on interval arithmetic, will be introduced. The use of interval-valued database indices will allow searching to accommodate scaling and geometric uncertainties without trial-and-error variation.
The system will be extended to humpback and sperm whale tail fluke images, introducing new features as indices.
Arrangements have been made for collaborative testing by prominent laboratories that maintain large collections of marine mammal sightings as photographic archives. The project will interact with these groups, supporting them as they use the project's system and using their criticisms and suggestions to improve the photo-identification system.
