Generative Adversarial Network Based Acoustic Scene Training Set Augmentation and Selection Using SVM Hyper-Plane
Generative Adversarial Network Based Acoustic Scene Training Set Augmentation and Selection Using SVM Hyper-Plane
复制标题
DOI：
--
复制
发表时间：
2017
期刊：
影响因子：
--
通讯作者：
Seongkyu Mun;Sangwook Park;D. Han;Hanseok Ko
中科院分区：
文献类型：
--
作者：
Seongkyu Mun;Sangwook Park;D. Han;Hanseok Ko
虽然通常预计使用大量标记的训练数据将导致深度学习的性能提高，但通常难以获得这样的数据库（DB）。在诸如声学场景和事件的检测和分类（DCASE）挑战任务1的比赛中，参与者被约束为使用相对小的DB作为规则，这类似于上述问题。为了在不增加训练数据库的情况下提高声场景分类的性能，提出了一种基于生成对抗网络（GAN）的方法来生成额外的训练数据库。由于不清楚GAN生成的每个样本是否对分类性能具有相同的影响，本文提出使用每个类别的支持向量机（SVM）超平面作为选择样本的参考，这些样本具有类别区分信息。基于开发数据库的交叉验证实验表明，使用生成的特征可以提高ASC的性能。
Although it is typically expected that using a large amount of labeled training data would lead to improve performance in deep learning, it is generally difficult to obtain such DataBase (DB). In competitions such as the Detection and Classification of Acoustic Scenes and Events (DCASE) challenge Task 1, participants are constrained to use a relatively small DB as a rule, which is similar to the aforementioned issue. To improve Acoustic Scene Classification (ASC) performance without employing additional DB, this paper proposes to use Generative Adversarial Networks (GAN) based method for generating additional training DB. Since it is not clear whether every sample generated by GAN would have equal impact in classification performance, this paper proposes to use Support Vector Machine (SVM) hyper plane for each class as reference for selecting samples, which have class discriminative information. Based on the crossvalidated experiments on development DB, the usage of the generated features could improve ASC performance.