Abstract
This paper addresses the problem of identifying individual animals in images based on extracting and matching contours, focusing in particular on the trailing edges of humpback whale flukes and the outline of the ears of African savanna elephants. A coarse-grained FCNN is learned to isolate the contour in an image, and a fine-grained FCNN is learned to provide more precise boundary information. The latter is trained by generating synthetic boundaries from coarse, easily-extracted training data, avoiding tedious manual effort. An A* algorithm extracts the final contour, which is converted to set of digital curvature descriptors and matched against a database of descriptors using local-naive Bayes nearest neighbors. We show that using the learned fine-grained FCNN produces more accurate contours than using image gradients for fine localization, especially for elephant ears where the boundaries are primarily texture. Matching using contours extracted using the fine-grained FCNN improves top-1 accuracy from 80% to 85% for flukes and 78% to 84% for ears.
Introduction
Conservation biology relies heavily on "photo-identification"—the ability to track individual animals over time using natural markings like spots, stripes, or notches. For species like humpback whales and African elephants, the "trailing edge" of a fluke or the "outer rim" of an ear acts like a fingerprint.
However, automating this is difficult. Traditional computer vision relies on intensity gradients to find edges. In the wild, these methods are often distracted by "noise" like ocean waves, forest shadows, or skin scars. Furthermore, some identifying features—especially on elephant ears—are subtle changes in texture rather than sharp edges. Training deep learning models usually requires thousands of pixel-perfect manual outlines, creating a significant bottleneck for field researchers.
Methodology
The proposed multi-stage pipeline balances human effort with machine precision through a two-stage neural network approach:
- Coarse-Grained FCNN (CG-FCNN): Annotators provide "thick brush" strokes that roughly cover the contour. The CG-FCNN learns to identify the general region of interest, suppressing background distractions.
- Self-Supervised Fine-Grained FCNN (FG-FCNN): Instead of pixel-perfect lines, the system uses coarse data to generate synthetic boundaries. It blends image patches from either side of the coarse line using random polynomial curves to create realistic training examples. The FG-FCNN learns to find the exact boundary even when visual transitions are subtle.
- Contour Extraction via A* Search: The outputs from both networks are combined into a cost matrix. An A* pathfinding algorithm extracts a single, continuous identifying contour.
- Matching: The extracted contour is converted into "integral curvature descriptors" and matched against a database using a Local Naive Bayes Nearest Neighbor (LNBNN) classifier.
Results and Discussion
The learned appearance model consistently outperformed traditional gradient-based methods across two primary real-world datasets: Humpback Whales (6,912 encounters) and African Savanna Elephants (508 encounters).
- Humpback Whales: Top-1 identification accuracy increased from 80% to 85%.
- African Elephants: Top-1 identification accuracy jumped from 78% to 84%.
The model excels at identifying elephant ears where the "edge" is often just a slight change in skin texture. By training a model to recognize what a fluke or ear looks like rather than just looking for sharp lines, the system ignores distracting high-contrast edges like the horizon or tree branches.
Conclusion
While elephant ear recognition remains more challenging than whale fluke recognition due to more localized identifying information, the results are promising. Future work may move toward a multi-modal approach—combining ear contours with other features like tusk shape or facial wrinkles. Currently, the system serves as a powerful tool for augmentation, providing a rank-ordered list for human experts to make final identifications.
Summary updated