Low-Resolution Image Classification using Knowledge Distillation from High-Resolution Image via Self-Attention Map

Abstract

With the number of media growing rapidly, consumers try to choose more attractive one from the floods of programs. Following these trends, many providers and developers focus on delivering realistic experiences using haptic feedback, visual information and etc. with their program. Our paper proposes vibro-acoustical feedback for the reality in game playing by catching the certain effect sound. The game sound of “Battle Ground” is used for this research, because it contains many effects that would be more realistic if more feedbacks were given. Dataset is collected from the Youtube and annotated manually. Only applying the volume filters, event sound, such as gun and bomb, could be detected well. However not only the event sound takes large portion of the data, but also the voice, chat by games users, takes. Because the volume of voice is usually high for clear communications, it doesn’t filtered by volume. So we add simple convolutional neural networks (CNN) for classifying the voice and event sound after volume filters. And reaches 96.32% accuracy. Our networks also evaluated with external data in real-worlds, and show the acceptable results. It means simulated sound data is applicable to real-world sound.