| Presentation | 2022-05-13 A serial anomalous sound detection method using outlier exposure based on two types of binary classification Ibuki Kuroyanagi, Tomoki Hayashi, Kazuya Takeda, Tomoki Toda, |
|---|---|
| PDF Download Page |  PDF download Page Link |
| Abstract(in Japanese) | (See Japanese page) |
| Abstract(in English) | Anomalous sound detection systems use only normal sound data to detect unknown, atypical sounds. Conventional methods use a serial method, a combination of outlier exposure, which classifies normal and pseudo-anomalous data and obtains embedding, and inlier modeling, which models the probability distribution of the embedding. Outlier exposure has a difficulty in training a good classifier when normal data and pseudo-anomalous data are too similar or too different. To explicitly distinguish cases where normal data and pseudo-anomalous data are too similar or too different, the proposed method performs two types of binary classification tasks when training outlier exposure. It allows more anomalous data to be detected. Evaluation results on the DCASE~2021 Task~2 dataset show that the proposed method, with a single model, outperforms the top methods that ensemble multiple models by 2.1,% in the harmonic mean of AUC and pAUC ($p=0.1$). |
| Keyword(in Japanese) | (See Japanese page) |
| Keyword(in English) | anomalous sound detection / outlier exposure / inlier modeling / hypersphere / multi-task learning |
| Paper # | EA2022-8 |
| Date of Issue | 2022-05-06 (EA) |
| Conference Information | |
| Committee | EA |
|---|---|
| Conference Date | 2022/5/13(1days) |
| Place (in Japanese) | (See Japanese page) |
| Place (in English) | Online |
| Topics (in Japanese) | (See Japanese page) |
| Topics (in English) | |
| Chair | Yoshinobu Kajikawa(Kansai Univ.) |
| Vice Chair | Kenichi Furuya(Oita Univ.) / Shoichi Koyama(Univ. of Tokyo) |
| Secretary | Kenichi Furuya(NTT) / Shoichi Koyama(RitsumeikanUniv.) |
| Assistant | Yukou Wakabayashi(Tokyo Metropolitan Univ.) / Tatsuya Komatsu(LINE) |
| Paper Information | |
| Registration To | Technical Committee on Engineering Acoustics |
|---|---|
| Language | JPN |
| Title (in Japanese) | (See Japanese page) |
| Sub Title (in Japanese) | (See Japanese page) |
| Title (in English) | A serial anomalous sound detection method using outlier exposure based on two types of binary classification |
| Sub Title (in English) | |
| Keyword(1) | anomalous sound detection |
| Keyword(2) | outlier exposure |
| Keyword(3) | inlier modeling |
| Keyword(4) | hypersphere |
| Keyword(5) | multi-task learning |
| 1st Author's Name | Ibuki Kuroyanagi |
| 1st Author's Affiliation | Nagoya University(Nagoya Univ.) |
| 2nd Author's Name | Tomoki Hayashi |
| 2nd Author's Affiliation | Nagoya University/Human Dataware Lab. Co. Ltd.(Nagoya Univ./HDL/) |
| 3rd Author's Name | Kazuya Takeda |
| 3rd Author's Affiliation | Nagoya University(Nagoya Univ.) |
| 4th Author's Name | Tomoki Toda |
| 4th Author's Affiliation | Nagoya University(Nagoya Univ.) |
| Date | 2022-05-13 |
| Paper # | EA2022-8 |
| Volume (vol) | vol.122 |
| Number (no) | EA-20 |
| Page | pp.pp.35-40(EA), |
| #Pages | 6 |
| Date of Issue | 2022-05-06 (EA) |