Two-Stage Classification Learning for Open Set Acoustic Scene Classification

摘要

Most of the research on acoustic scene classification (ASC) focuses on classification problem with only known scene classes. In practice, scene classification problem to be solved generally is based on an open set, which contains unknown scenes. This paper proposes a two-stage method that solves the open set problem on ASC. The proposed system decomposes open set ASC problem into two stages. To mitigate the impact of unknown scenes on the subsequent recognition process of known scenes, the first stage is to identify unknown scenes. The second stage classifies defined acoustic scenes. In this case, the threshold selection strategy we proposed further sorts out unknown scenes that were not identified in the previous stage. Experiments show that the method proposed in this paper can effectively identify unknown scenes and classify known scenes, by segmenting the open set acoustic scene classification task and selecting an appropriate judgment threshold. On the development dataset released by DCASE Challenge 2019 Task 1C, the model proposed outperforms the first place.