Dynamic Neural Network Optimization Framework for Adaptive Sensor Selection in Depth Imaging and Registration

Abstract

Accurate and efficient sensor selection is a cornerstone for robust 2D and 3D depth imaging and registration, with applications spanning autonomous vehicles, robotics, and augmented reality systems. Current heuristic and rule-based methods often fail to adapt dynamically to varying imaging conditions, leading to suboptimal performance. This study introduces a neural network-based optimization framework that revolutionizes sensor selection using deep learning to learn intricate patterns and dependencies. Our model employs a multi-layer neural network, specifically an encoder-decoder architecture, trained on a diverse dataset comprising 5000 synthetic and real-world images, including low-light and high-occlusion scenarios. The model was trained using the Adam optimizer with a learning rate of 0.001. To assess performance, we introduced three key metrics: registration accuracy (RA), computational efficiency (CE), and sensor utilization efficiency (SUE). The proposed framework outperformed benchmark models, achieving a +28.7% ± 1.8 improvement in RA, a +32.4% ± 2.1 increase in CE, and a +26.3% ± 1.5 enhancement in SUE compared to ResNet-50 and EfficientNet-B3 models. Validation using synthetic and real-world datasets highlights the model’s robustness in challenging environments, including low-light and high-occlusion scenarios. Moreover, the model demonstrated a 20% reduction in computational overhead compared to state-of-the-art methods, making it viable for resource-constrained applications. This research establishes a scalable and adaptive solution for sensor optimization, setting a new benchmark in depth imaging and registration