Energy-Efficient Saliency-Guided Video Coding Framework for Real-Time Applications

Abstract

The significant growth in global video data traffic can be mitigated by saliency-based video coding schemes that seek to increase coding efficiency without any loss of objective visual quality by compressing salient video regions less heavily than non-salient regions. However, conducting salient object detection (SOD) on every video frame before encoding tends to lead to substantial complexity and energy consumption overhead, especially if state-of-the-art deep learning techniques are used in saliency detection. This work introduces a saliency-guided video encoding framework that reduces the energy consumption over frame-by-frame SOD by increasing the detection interval and applying the proposed region-of-interest (ROI) tracking between successive detections. The computational complexity of our ROI tracking technique is kept low by predicting object movements from motion vectors, which are inherently calculated during encoding. Our experimental results demonstrate that the proposed ROI tracking solution saves energy by 86-95% and attains 84-94% accuracy over frame-by-frame SOD. Correspondingly, integrating our proposal into the complete saliency-guided video coding scheme reduces energy consumption on CPU by 79-82% at a cost of weighted PSNR of less than 5%. These findings indicate that our solution has significant potential for low-cost and low-power streaming media applications.