Model Compression via Structural Pruning and Feature Distillation for Accurate Multi-Spectral Object Detection on Edge-Devices

Multi-spectral infrared object detection across different infrared wavelengths is a challenging task. Although some full-sized object detection models, such as YOLOv4 and ScaledYOLO, may achieve good infrared object detection, they are resource-demanding and unsuitable for real-time detection on edge devices. Tiny versions for object detection are proposed to meet the practical requirement, but they usually sacrifice model accuracy and generalization for efficiency. We propose an accurate and efficient object detector capable of performing real-time inference under the hardware constraints of an edge device by leveraging structural pruning, feature distillation, and neural architecture search (NAS). The experiments on FLIR and multi-spectral object detection datasets show that our model achieves comparable mAP to full-sized models while having 14x times fewer parameters and 3.5x times fewer FLOPs. Our model can perform infrared detection well across different infrared wavelengths. The optimal CSPNet configurations of our detection network selected by NAS show that the resulting architectures outperform the baseline.