Neural style transfer combined with EfficientDet for thermal surveillance

Hindrance caused while performing object detection during border or perimeter surveillance at night due to low resolution and atmospheric noise inside infrared frames, demands for a sophisticated detection framework. To deal with the challenges, we present a novel deep learning framework dedicated completely for automated thermal surveillance. The framework is a composition of two major modules. A novel style transferred enhanced image module STEIM enhances the resolution of IR input frame by utilizing the contextual information and exploiting the local and global features to preserve the high frequency details. The enhanced IR frame is further fed inside the fine-tuned EfficientDet module EDM that comprises weighted two-way feature network. It has an advantage of effectual multi-scale feature fusion contributing to accurate and efficient object detection. We have obtained mAP of 92.83% with 94.07% accuracy running at 95fps on FLIR test frames and mAP of 87.51% with 88.53% accuracy running at 89fps on OTCVBS test frames. Experimental results on two benchmark datasets FLIR and OTCBVS completely demonstrate suitability of the proposed framework for automatic surveillance using thermal imaging systems.