Neural Network-Based Sensor-Guided Exposure Control Eliminating Multi-Frame Input

Camera-based IoT nodes require efficient duty cycling to maximize battery life and minimize maintenance needs. However, existing exposure control methods require complex optimization, multi-frame convergence, and high computational resources, resulting in a long latency to capture a single image. In this paper, we propose a camera exposure control method using a hybrid approach based on a neural network model to predict camera exposure time settings. The model uses light sensor values, image gradients, entropy, and overexposure characteristics as input. A customized dataset collection strategy is designed to train the model under different conditions. Experimental test results show an exposure time skew of 1.12 ms compared to the reference without the need to capture images for control, making the proposed method suitable for IoT-based smart cameras.