AAI_2025_Capstone_Chronicles_Combined

Figure 16. WildScan Demo

By making the entire process transparent and interactive, the demo bridges the gap between technical model performance and real-world usability. It not only showcases our work but also actively collects high-quality human feedback to refine the model in future iterations. The development of ScratchResNet demonstrated that a fully custom convolutional neural network, trained entirely from scratch, can achieve competitive performance in wildlife image classification. Achieving ~74% validation accuracy without pre-trained weights validates both the architecture design and the robustness of the training pipeline. The normalized confusion matrix revealed strong class-level performance for several species, while also highlighting specific areas of confusion — particularly among visually similar classes and background detections.

Building on these results, future work will focus on several areas. First, we will enhance the architecture by integrating attention modules such as Squeeze-and-Excitation and CBAM and selectively deepening

298