Improving Object Counting Accuracy with Adaptive CNN Models and Meta-Level Routing
Benjamin Bowman, Zhi Zheng, Nesreen Dalhy, Brendan Geary, Ian Bentley, Bayazit Karaman · 2025
Accurate estimation of bat populations is essential for ecological monitoring, disease prevention, and conservation planning, yet traditional manual counting techniques remain inefficient and error-prone. We propose an automated video-based bat counting framework that combines temporal median background subtraction with a hybrid deep learning architecture. The system extracts regions of interest through background subtraction and classifies bat counts using a generalized Convolutional Neural Network (CNN) integrated with a Mixture of Experts (MoE) routing mechanism, which dynamically assigns images to specialized CNNs based on inferred bat density levels. To address class imbalance, particularly in high-density scenarios, we introduce an adaptive synthetic data generation method that augments the training dataset. Experimental results show a 7% improvement in classification accuracy over a traditional CNN baseline, with consistent performance across varying density conditions. Additionally, our approach improves classification accuracy on synthetic bat images from prior work, increasing it from 63% to 71%. This framework offers a scalable, noninvasive, and effective solution for real-world bat population monitoring, providing valuable support for ecological research and conservation efforts.