針對圖形密集型應用之輕量級影像相似度評估

Abstract

Content similarity measurement enables mobile applications to reduce unnecessary computation energy without compromising user visual experience. Existing measures assess the similarity between image frames based on pixels, resulting in non-negligible overhead. This paper presents a lightweight similarity measure called LSIM, which assesses content similarity based on objects in graphics applications. To evaluate the efficacy, we implement LSIM in Android and conduct extensive experiments on a commercial smartphone with various mobile gaming applications. The results show that LSIM is highly correlated with a commonly used pixel-wise measure while incurring nearly zero overhead. We also apply LSIM to a CPU-GPU governing framework to mitigate the rendering of similar frames, thereby reducing the energy consumption by up to 27.3% while maintaining satisfactory visual quality.