Sea turtle carapace pattern as criterion for automatic species ID

The pattern seen on the carapace of sea turtles is one of the criterion used for identifying species, why not try to use this for automatic species ID as well ? The advantages of using this approach for automatic ID would be

1. Works for photos taken in wide range of environments and lighting condition varying from the animal's natural habitat (under water) to decaying carcass of stranded animals found on beaches
2. Only the carapace needs to be observed; other body parts, which decay sooner than the carapace in a dead animal, don't need to be observed
   
3. Can work for both color and bw photographs
4. Possibility of successful detection from broken or incomplete carapace if the key parts of the carapace is found

However, the challenges associated with this approach are also great. The challenges are encountered in two different phases, each of them are explained bellow.

• Feature Extraction
  The sought after feature here is the pattern on the carapace. To understand the challenges in extracting this feature, we should understand its origin first. What we are referring to as pattern in common English as 'Pattern' is actually the network of ridge lines between scutes on the turtles carapace. This is true for all the species of Cheloniida family. For Dermochelys coriacea or leatherback turtle, the only species of Dermochelyidae family, the pattern is originated by the seven ridge lines on its back. Experimentation with several photograph taken in varying environment reveals that the ridge lines are best visible, hence extractable, in case of underwater photos.
  
  
  
  The above image shows the result of gradient magnitude, a standard edge detection algorithm, applied on a photograph followed by edge thinning on the resulting gradient magnitude image. But this does not yield good extractable features from photographs of stranded animals found on beaches, specially the decaying ones. Worse, due to presence of debris or other objects inside the frame, wrong part of the image is extracted instead of the carapace by the gradient magnitude finding operation.
  
  




• Graph reconstruction
  The next challenge encountered is reconstruction of the pattern extracted from the photograph. Ideally, this would be the complete network of ridge lines. But the resulting image after feature extraction does not always have a complete network traced in most of the real life situations. So the graph reconstruction procedure should be robust enough to reconstruct a graph from incomplete or broken edges and at the same time be able to identify and discard edges that are not from the ridge lines on the turtles back.