AAI_2025_Capstone_Chronicles_Combined

The normalized confusion matrix (Figure 8) provides deeper insight into per-class performance, revealing clear strengths and well-defined limitations. Species such as cat, coyote, and raccoon were identified with high precision, suggesting the network learned highly distinctive and robust visual cues for these categories. However, certain visually or contextually similar species, including fox and coyote or skunk and raccoon, exhibited higher confusion rates. These errors are likely attributable to overlapping fur patterns, body shapes, and environmental contexts, which can make them challenging to separate even for human observers. What is most encouraging is that the model’s errors are not random; rather, they tend to cluster around semantically and visually related categories, indicating that ScratchResNet has learned meaningful decision boundaries. This suggests that targeted refinements could yield significant improvements. In particular, data augmentation focused on the most ambiguous species could help the network differentiate subtle visual cues. Likewise, architectural enhancements such as attention mechanisms could enable the model to focus more selectively on informative regions, while class-specific hard negative mining could

290