Fig. 2
An overview of the proposed methodology, which improves upon the accuracy of a single global detection model trained on all animal classes through a multi-tier expert system. Differentiation Between Similar Species: In scenario A, a camera trap captures an image mistakenly identified by the global model as a hare (label 20) with 85% confidence. However, this image contains a rabbit. The global model’s initial output is then processed by an expert model specializing in smaller mammals, correctly identifying the rabbit (label 21) with a higher confidence of 96%. This demonstrates our specialized expert models’ effectiveness in refining the global model’s initial predictions, especially in distinguishing closely related species. Detection of Animals in Empty-Looking Images: In scenario B, the global model fails to detect a small, hidden animal (a mouse) and classifies the image as “background” (label 24). Then, the response is redirected to all expert models to re-evaluate the image. Each expert model provides a decision, and the final identification is determined through a voting system based on confidence levels. Ultimately, the expert model for micro-mammals correctly identifies the mouse with a confidence of 91%, showcasing our method’s robustness in detecting animals even in challenging, low-visibility conditions.
