AAI_2025_Capstone_Chronicles_Combined

for deeper models. Unlike conventional ResNet variants that initialize from large pre-trained backbones (e.g., ImageNet), ScratchResNet was initialized entirely from scratch, allowing the network to learn feature representations directly from the wildlife dataset rather than adapting from unrelated domains.

Data Preparation & Baseline Model​ ​ Data preparation involved parsing CSV annotations to accurately map each cropped image to its corresponding species label. This preprocessing pipeline ensured consistent image–label alignment across the dataset. Before committing to the custom architecture, a simple CNN baseline was trained with standard augmentations (random flips, rotations, normalization). This baseline achieved ~30% validation accuracy — enough to confirm that the preprocessing and labeling pipeline were functioning correctly. More importantly, the baseline results highlighted areas where a deeper, more expressive architecture could yield substantial performance gains. We maximized model performance by leveraging Optuna for Bayesian hyperparameter optimization, systematically tuning six key parameters—namely, the number of residual blocks per stage, channel width multipliers, dropout rates, initial learning rate and scheduling parameters, weight decay, and batch size—and selecting the optimal configuration based on validation accuracy for final training. To further bolster generalization, we devised a comprehensive augmentation pipeline comprising MixUp (α = 0.4) to

288