H-BNN: FPGA-based binarized convolutional neural network for cloud detection on satellite payload

Abstract

This paper proposes an FPGA-based binarized convolutional neural network (H-BNN) for cloud detection on a satellite payload. By utilizing 1-bitwise weights and activations, the proposed approach reduces computational complexity and memory requirements, making it an efficient solution for classifying cloud regions in near-infrared images captured by satellite camera sensors. The proposed H-BNN model and hardware implementation approach are more suitable for satellite (a) payload hardware computing and address the challenges posed by traditional Convolution Neural Network models with full precision configuration., and achieved an accuracy of over 90% and 22 Frames Per Second(FPS) on FPGA.