음향 이벤트 검출을 위한 DenseNet-Recurrent Neural Network 학습 방법에 관한 연구 = A study on training DenseNet-Recurrent Neural Network for sound event detection

국문 초록 (Abstract)

음향 이벤트 검출(Sound Event Detection, SED)은 음향 신호에서 관심 있는 음향의 종류와 발생 구간을검출하는 기술로, 음향 감시 시스템 및 모니터링 시스템 등 다양한 분야에서 활용되고 있다. 최근 음향 신호 분석에 관한국제 경연 대회(Detection and Classification of Acoustic Scenes and Events, DCASE) Task 4를 통해 다양한 방법이소개되고 있다. 본 연구는 다양한 영역에서 성능 향상을 이끌고 있는 Dense Convolutional Networks(DenseNet)을음향 이벤트 검출에 적용하기 위해 설계 변수에 따른 성능 변화를 비교 및 분석한다. 실험에서는 DenseNet with Bottleneck and Compression(DenseNet-BC)와 순환신경망(Recurrent Neural Network, RNN)의 한 종류인 양방향게이트 순환 유닛(Bidirectional Gated Recurrent Unit, Bi-GRU)을 결합한 DenseRNN 모델을 설계하고, 평균 교사모델(Mean Teacher Model)을 통해 모델을 학습한다. DCASE task4의 성능 평가 기준에 따라 이벤트 기반 f-score를바탕으로 설계 변수에 따른 DenseRNN의 성능 변화를 분석한다. 실험 결과에서 DenseRNN의 복잡도가 높을수록 성능이 향상되지만 일정 수준에 도달하면 유사한 성능을 보임을 확인할 수 있다. 또한, 학습과정에서 중도탈락을 적용하지 않는 경우, 모델이 효과적으로 학습됨을 확인할 수 있다.

다국어 초록 (Multilingual Abstract)

Sound Event Detection (SED) aims to identify not only sound category but also time interval for target sounds in an audio waveform. It is a critical technique in field of acoustic surveillance system and monitoring system. Recently, various models have introduced through Detection and Classification of Acoustic Scenes and Events (DCASE) Task 4. This paper explored how to design optimal parameters of DenseNet based model, which has led to outstanding performance in other recognition system. In experiment, DenseRNN as an SED model consists of DensNet-BC and bi-directional Gated Recurrent Units (GRU). This model is trained with Mean teacher model. With an event-based f-score, evaluation is performed depending on parameters, related to model architecture as well as model training, under the assessment protocol of DCASE task4. Experimental result shows that the performance goes up and has been saturated to near the best. Also, DenseRNN would be trained more effectively without dropout technique.

참고문헌 (Reference)

1 S. Zagoruyko, "Wide residual networks"

2 X. Glorot, "Understanding the difficulty of training deep feedforward neural networks" 249-256, 2010

3 S. Park, "Time-balanced focal loss for audio event detection" 311-315, 2022

4 N. Turpault, "Sound event detection in domestic environments with weakly labeled data and soundscape synthesis" 253-257, 2019

5 A. Mesaros, "Sound event detection : A tutorial" 38 : 67-83, 2021

6 K. Palanisamy, "Rethinking CNN models for audio classification"

7 "PyTorch Torch.nn.GRU"

8 A. Tarvainen, "Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results" 1-10, 2017

9 L. Delphin-Poulat, "Mean teacher with data augmentation for dcase 2019 task 4" Orange Labs Lannion 2019

10 B. McMahan, "Listening to the world improves speech command recognition" 378-385, 2018

11 K. He, "Identity mappings in deep residual networks" 1-15, 2016

12 N. Srivastava, "Dropout: a simple way to prevent neural networks from overfitting" 15 : 2014

13 G. Huang, "Densely connected convolutional networks" 1-9, 2017

14 K. He, "Delving deep into rectifiers : Surpassing human-level performance on imagenet classification" 1026-1032, 2015

15 X. Glorot, "Deep sparse rectifier neural networks" 315-323, 2011

16 K. He, "Deep residual learning for image recognition" 770-778, 2016

17 "DCASE 2020 Task 4: Sound Event Detection and Separation in Domestic Environments"

18 "DCASE 2020 Task 4 GitHub"

19 S. Ioffe, "Batch normalization: Accelerating deep network training by reducing internal covariate shift"

20 D. P. Kingma, "Adam: A method for stochastic optimization"

21 "A PyTorch Implementation for Densely Connected Convolutional Networks (DenseNets)"