Pub Date : 2021-01-24DOI: 10.23919/Eusipco47968.2020.9287538
A. Lawal, Qadri Mayyala, A. Zerguine, Azeddine Beghdadi
This work presents a method for salt dome detection in seismic images based on a Context-Aware Saliency (CAS) detection model. Seismic data can easily add up to hundred of gigabytes and terabytes in size. However, the key features or structural information that are of interest to the seismic interpreters are quite few. These features include salt domes, fault and other geological features that have the potential of indicating the presence of oil reservoir. A new method for extracting the most perceptual relevant features in seismic images based on the CAS model is proposed. The efficiency of this method in detecting the most salient structures in a seismic image such as salt dome is demonstrated through a series of experiment on real data set with various spatial contents.
{"title":"Salt Dome Detection Using Context-Aware Saliency","authors":"A. Lawal, Qadri Mayyala, A. Zerguine, Azeddine Beghdadi","doi":"10.23919/Eusipco47968.2020.9287538","DOIUrl":"https://doi.org/10.23919/Eusipco47968.2020.9287538","url":null,"abstract":"This work presents a method for salt dome detection in seismic images based on a Context-Aware Saliency (CAS) detection model. Seismic data can easily add up to hundred of gigabytes and terabytes in size. However, the key features or structural information that are of interest to the seismic interpreters are quite few. These features include salt domes, fault and other geological features that have the potential of indicating the presence of oil reservoir. A new method for extracting the most perceptual relevant features in seismic images based on the CAS model is proposed. The efficiency of this method in detecting the most salient structures in a seismic image such as salt dome is demonstrated through a series of experiment on real data set with various spatial contents.","PeriodicalId":6705,"journal":{"name":"2020 28th European Signal Processing Conference (EUSIPCO)","volume":"21 1","pages":"1906-1910"},"PeriodicalIF":0.0,"publicationDate":"2021-01-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"89873611","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-01-24DOI: 10.23919/Eusipco47968.2020.9287435
Tayebeh Lotfi Mahyari, R. Dansereau
One way to solve under-determined image decomposition is to use statistical information about the type of data to be decomposed. This information can be obtained by a deep learning where convolutional neural networks (CNN) are a subset recently used widely in image processing. In this paper, we have designed a two-stage CNN that takes cytology images of overlapped cervical cells and attempts to separate the cell images. In the first stage, we designed a CNN to segment overlapping cells. In the second stage, we designed a CNN that uses this segmentation and the original image to separate the regions. We implemented a CNN similar to U-Net for image segmentation and implemented a new network for the image separation. To train and test the proposed networks, we simulated 50000 cervical cell cytology images by overlaying individual images of real cervical cells using the Beer-Lambert law. Of these 50000 images, we used 49000 images for training and evaluated the method with 1000 test images. Results on these synthetic images give more than 97% segmentation accuracy and gives decomposition SSIM scores of more than 0.99 and PSNR score of more than 30 dB. Despite these positive results, the permutation problem that commonly effects signal separation occasionally occurred resulting in some cell structure mis-separation (for example, one cell given two nucleoli and the other given none). In addition, when the segmentation was poor from the first stage, the resulting separation was poor.
{"title":"Deep Learning Methods for Image Decomposition of Cervical Cells","authors":"Tayebeh Lotfi Mahyari, R. Dansereau","doi":"10.23919/Eusipco47968.2020.9287435","DOIUrl":"https://doi.org/10.23919/Eusipco47968.2020.9287435","url":null,"abstract":"One way to solve under-determined image decomposition is to use statistical information about the type of data to be decomposed. This information can be obtained by a deep learning where convolutional neural networks (CNN) are a subset recently used widely in image processing. In this paper, we have designed a two-stage CNN that takes cytology images of overlapped cervical cells and attempts to separate the cell images. In the first stage, we designed a CNN to segment overlapping cells. In the second stage, we designed a CNN that uses this segmentation and the original image to separate the regions. We implemented a CNN similar to U-Net for image segmentation and implemented a new network for the image separation. To train and test the proposed networks, we simulated 50000 cervical cell cytology images by overlaying individual images of real cervical cells using the Beer-Lambert law. Of these 50000 images, we used 49000 images for training and evaluated the method with 1000 test images. Results on these synthetic images give more than 97% segmentation accuracy and gives decomposition SSIM scores of more than 0.99 and PSNR score of more than 30 dB. Despite these positive results, the permutation problem that commonly effects signal separation occasionally occurred resulting in some cell structure mis-separation (for example, one cell given two nucleoli and the other given none). In addition, when the segmentation was poor from the first stage, the resulting separation was poor.","PeriodicalId":6705,"journal":{"name":"2020 28th European Signal Processing Conference (EUSIPCO)","volume":"52 1","pages":"1110-1114"},"PeriodicalIF":0.0,"publicationDate":"2021-01-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"90549014","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-01-24DOI: 10.23919/Eusipco47968.2020.9287753
Takuya Hasumi, Tetsunori Kobayashi, Tetsuji Ogawa
This paper examines the effectiveness of a fully convolutional time-domain audio separation network (Conv-TasNet) on single-channel denoising. Conv-TasNet, which has a structure to explicitly estimate a mask for encoded features, has shown to be effective in single-channel sound source separation in noise-free environments, but it has not been applied to denoising. Therefore, the present study investigates a method of learning Conv-TasNet for denoising and clarifies the optimal structure for single-channel end-to-end modeling. Experimental comparisons conducted using the CHiME-3 dataset demonstrate that Conv-TasNet performs well in denoising and yields improvements in single-channel end-to-end denoising over existing denoising autoencoder-based modeling.
{"title":"Investigation of Network Architecture for Single-Channel End-to-End Denoising","authors":"Takuya Hasumi, Tetsunori Kobayashi, Tetsuji Ogawa","doi":"10.23919/Eusipco47968.2020.9287753","DOIUrl":"https://doi.org/10.23919/Eusipco47968.2020.9287753","url":null,"abstract":"This paper examines the effectiveness of a fully convolutional time-domain audio separation network (Conv-TasNet) on single-channel denoising. Conv-TasNet, which has a structure to explicitly estimate a mask for encoded features, has shown to be effective in single-channel sound source separation in noise-free environments, but it has not been applied to denoising. Therefore, the present study investigates a method of learning Conv-TasNet for denoising and clarifies the optimal structure for single-channel end-to-end modeling. Experimental comparisons conducted using the CHiME-3 dataset demonstrate that Conv-TasNet performs well in denoising and yields improvements in single-channel end-to-end denoising over existing denoising autoencoder-based modeling.","PeriodicalId":6705,"journal":{"name":"2020 28th European Signal Processing Conference (EUSIPCO)","volume":"4 1","pages":"441-445"},"PeriodicalIF":0.0,"publicationDate":"2021-01-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"91433906","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-01-24DOI: 10.23919/Eusipco47968.2020.9287730
Alexander Bartler, Darius Hinderer, Bin Yang
Nowadays, the explainability of deep neural networks is an essential part of machine learning. In the last years, many methods were developed to visualize important regions of an input image for the decision of the deep neural network. Since almost all methods are designed for supervised trained models, we propose in this work a visualization technique for unsupervised trained autoencoders called Gradient-weighted Latent Activation Mapping (Grad-LAM). We adapt the idea of Grad-CAM and propose a novel weighting based on the knowledge of the autoencoder’s decoder. Our method will help to get insights into the highly nonlinear mapping of an input image to a latent space. We show that the visualization maps of Grad-LAM are meaningful on simple datasets like MNIST and the method is even applicable to real-world datasets like ImageNet.
{"title":"Grad-LAM: Visualization of Deep Neural Networks for Unsupervised Learning","authors":"Alexander Bartler, Darius Hinderer, Bin Yang","doi":"10.23919/Eusipco47968.2020.9287730","DOIUrl":"https://doi.org/10.23919/Eusipco47968.2020.9287730","url":null,"abstract":"Nowadays, the explainability of deep neural networks is an essential part of machine learning. In the last years, many methods were developed to visualize important regions of an input image for the decision of the deep neural network. Since almost all methods are designed for supervised trained models, we propose in this work a visualization technique for unsupervised trained autoencoders called Gradient-weighted Latent Activation Mapping (Grad-LAM). We adapt the idea of Grad-CAM and propose a novel weighting based on the knowledge of the autoencoder’s decoder. Our method will help to get insights into the highly nonlinear mapping of an input image to a latent space. We show that the visualization maps of Grad-LAM are meaningful on simple datasets like MNIST and the method is even applicable to real-world datasets like ImageNet.","PeriodicalId":6705,"journal":{"name":"2020 28th European Signal Processing Conference (EUSIPCO)","volume":"68 1","pages":"1407-1411"},"PeriodicalIF":0.0,"publicationDate":"2021-01-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"91347886","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-01-24DOI: 10.23919/Eusipco47968.2020.9287568
G. Tzagkarakis, F. Maurer, T. Dionysopoulos
Quantifying risk is pivotal for every financial institution, with the temporal dimension being the key aspect for all the well-established risk measures. However, exploiting the frequency information conveyed by financial data, could yield improved insights about the inherent risk evolution in a joint time-frequency fashion. Nevertheless, the great majority of risk managers make no explicit distinction between the information captured by patterns of different frequency content, while relying on the full time-resolution data, regardless of the trading horizon. To address this problem, a novel value-at-risk (VaR) quantification method is proposed, which combines nonlinearly the time-evolving energy profile of returns series at multiple frequency scales, determined by the predefined trading horizon. Most importantly, our proposed method can be coupled with any quantile-based risk measure to enhance its performance. Experimental evaluation with real data reveals an increased robustness of our method in efficiently controlling under-/over-estimated VaR values.
{"title":"Nonparametric Adaptive Value-at-Risk Quantification Based on the Multiscale Energy Distribution of Asset Returns","authors":"G. Tzagkarakis, F. Maurer, T. Dionysopoulos","doi":"10.23919/Eusipco47968.2020.9287568","DOIUrl":"https://doi.org/10.23919/Eusipco47968.2020.9287568","url":null,"abstract":"Quantifying risk is pivotal for every financial institution, with the temporal dimension being the key aspect for all the well-established risk measures. However, exploiting the frequency information conveyed by financial data, could yield improved insights about the inherent risk evolution in a joint time-frequency fashion. Nevertheless, the great majority of risk managers make no explicit distinction between the information captured by patterns of different frequency content, while relying on the full time-resolution data, regardless of the trading horizon. To address this problem, a novel value-at-risk (VaR) quantification method is proposed, which combines nonlinearly the time-evolving energy profile of returns series at multiple frequency scales, determined by the predefined trading horizon. Most importantly, our proposed method can be coupled with any quantile-based risk measure to enhance its performance. Experimental evaluation with real data reveals an increased robustness of our method in efficiently controlling under-/over-estimated VaR values.","PeriodicalId":6705,"journal":{"name":"2020 28th European Signal Processing Conference (EUSIPCO)","volume":"15 1","pages":"2393-2397"},"PeriodicalIF":0.0,"publicationDate":"2021-01-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"82517821","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-01-24DOI: 10.23919/Eusipco47968.2020.9287664
M. Laakso, Robin Rajamäki, R. Wichman, V. Koivunen
We introduce a modular and affordable coherent multichannel software-defined radio (SDR) receiver and demonstrate its performance by direction-of-arrival (DOA) estimation on signals collected from a 7 X 3 element uniform rectangular array antenna, comparing the results between the full and sparse arrays. Sparse sensor arrays can reach the resolution of a fully populated array with reduced number of elements, which relaxes the required structural complexity of e.g. antenna arrays. Moreover, sparse arrays facilitate significant cost reduction since fewer expensive RF-IF front ends are needed. Results from the collected data set are analyzed with Multiple Signal Classification (MUSIC) DOA estimator. Generally, the sparse array estimates agree with the full array.
{"title":"Phase-coherent multichannel SDR - Sparse array beamforming","authors":"M. Laakso, Robin Rajamäki, R. Wichman, V. Koivunen","doi":"10.23919/Eusipco47968.2020.9287664","DOIUrl":"https://doi.org/10.23919/Eusipco47968.2020.9287664","url":null,"abstract":"We introduce a modular and affordable coherent multichannel software-defined radio (SDR) receiver and demonstrate its performance by direction-of-arrival (DOA) estimation on signals collected from a 7 X 3 element uniform rectangular array antenna, comparing the results between the full and sparse arrays. Sparse sensor arrays can reach the resolution of a fully populated array with reduced number of elements, which relaxes the required structural complexity of e.g. antenna arrays. Moreover, sparse arrays facilitate significant cost reduction since fewer expensive RF-IF front ends are needed. Results from the collected data set are analyzed with Multiple Signal Classification (MUSIC) DOA estimator. Generally, the sparse array estimates agree with the full array.","PeriodicalId":6705,"journal":{"name":"2020 28th European Signal Processing Conference (EUSIPCO)","volume":"31 1","pages":"1856-1860"},"PeriodicalIF":0.0,"publicationDate":"2021-01-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"78319010","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-01-24DOI: 10.23919/Eusipco47968.2020.9287322
Mariem Bouafif Mansali, Tomas Bäckström, Z. Lachiri
Multiple pitch streaming from a mixture is a challenging problem for signal processing and especially for speech separation. In this paper, we use a Zero frequency filtering (ZFF) based new system to stream pitch of multiple concurrent speakers. We propose a workflow to estimate pitch values of all sources in each single frame then streaming them into trajectories, each corresponding to a distinct source. The method consists of detecting and localizing the involved speakers in a mixture, followed by a ZFF based approach where involved speakers’ pitches are iteratively streamed from the observed mixture. The robustness of the proposed system is tested over two, and three overlapping speech mixtures collected in reverberant environment. The results indicate that our proposal brings ZFF to a competitive level with another recently proposed streaming approach.
{"title":"Evaluation of Zero Frequency Filtering based Method for Multi-pitch Streaming of Concurrent Speech Signals","authors":"Mariem Bouafif Mansali, Tomas Bäckström, Z. Lachiri","doi":"10.23919/Eusipco47968.2020.9287322","DOIUrl":"https://doi.org/10.23919/Eusipco47968.2020.9287322","url":null,"abstract":"Multiple pitch streaming from a mixture is a challenging problem for signal processing and especially for speech separation. In this paper, we use a Zero frequency filtering (ZFF) based new system to stream pitch of multiple concurrent speakers. We propose a workflow to estimate pitch values of all sources in each single frame then streaming them into trajectories, each corresponding to a distinct source. The method consists of detecting and localizing the involved speakers in a mixture, followed by a ZFF based approach where involved speakers’ pitches are iteratively streamed from the observed mixture. The robustness of the proposed system is tested over two, and three overlapping speech mixtures collected in reverberant environment. The results indicate that our proposal brings ZFF to a competitive level with another recently proposed streaming approach.","PeriodicalId":6705,"journal":{"name":"2020 28th European Signal Processing Conference (EUSIPCO)","volume":"5 1","pages":"286-290"},"PeriodicalIF":0.0,"publicationDate":"2021-01-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"78400381","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-01-24DOI: 10.23919/Eusipco47968.2020.9287610
Bianca Iancu, E. Isufi
This work proposes a learning framework for distributed finite-time consensus with graph convolutional neural networks (GCNNs). Consensus is a central problem in distributed and adaptive optimisation, signal processing, and control. We leverage the link between finite-time consensus and graph filters, and between graph filters and GCNNs to study the potential of a readily distributed architecture for reaching consensus. We have found GCNNs outperform classical graph filters for distributed consensus and generalize better to unseen topologies such as distributed networks affected by link losses.
{"title":"Towards Finite-Time Consensus with Graph Convolutional Neural Networks","authors":"Bianca Iancu, E. Isufi","doi":"10.23919/Eusipco47968.2020.9287610","DOIUrl":"https://doi.org/10.23919/Eusipco47968.2020.9287610","url":null,"abstract":"This work proposes a learning framework for distributed finite-time consensus with graph convolutional neural networks (GCNNs). Consensus is a central problem in distributed and adaptive optimisation, signal processing, and control. We leverage the link between finite-time consensus and graph filters, and between graph filters and GCNNs to study the potential of a readily distributed architecture for reaching consensus. We have found GCNNs outperform classical graph filters for distributed consensus and generalize better to unseen topologies such as distributed networks affected by link losses.","PeriodicalId":6705,"journal":{"name":"2020 28th European Signal Processing Conference (EUSIPCO)","volume":"22 1","pages":"2145-2149"},"PeriodicalIF":0.0,"publicationDate":"2021-01-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"78504572","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-01-24DOI: 10.23919/Eusipco47968.2020.9287326
M. Jaber, A. Nasser, N. Charara, A. Mansour, K. Yao
The main aim of the Spectrum Sensing (SS) in a Cognitive Radio system is to distinguish between the binary hypotheses H0: Primary User (PU) is absent and H1: PU is active. In this paper, Machine Learning (ML)-based hybrid Spectrum Sensing (SS) scheme is proposed. The scattering of the Test Statistics (TSs) of two detectors is used in the learning and prediction phases. As the SS decision is binary, the proposed scheme requires the learning of only the boundaries of H0-class in order to make a decision on the PU status: active or idle. Thus, a set of data generated under H0 hypothesis is used to train the detection system. Accordingly, unlike the existing ML-based schemes of the literature, no PU statistical parameters are required. In order to discriminate between H0-class and elsewhere, we used a one-class classification approach that is inspired by the Isolation Forest algorithm. Extensive simulations are done in order to investigate the efficiency of such hybrid SS and the impact of the novelty detection model parameters on the detection performance. Indeed, these simulations corroborate the efficiency of the proposed one-class learning of the hybrid SS system.
{"title":"One-Class based learning for Hybrid Spectrum Sensing in Cognitive Radio","authors":"M. Jaber, A. Nasser, N. Charara, A. Mansour, K. Yao","doi":"10.23919/Eusipco47968.2020.9287326","DOIUrl":"https://doi.org/10.23919/Eusipco47968.2020.9287326","url":null,"abstract":"The main aim of the Spectrum Sensing (SS) in a Cognitive Radio system is to distinguish between the binary hypotheses H0: Primary User (PU) is absent and H1: PU is active. In this paper, Machine Learning (ML)-based hybrid Spectrum Sensing (SS) scheme is proposed. The scattering of the Test Statistics (TSs) of two detectors is used in the learning and prediction phases. As the SS decision is binary, the proposed scheme requires the learning of only the boundaries of H0-class in order to make a decision on the PU status: active or idle. Thus, a set of data generated under H0 hypothesis is used to train the detection system. Accordingly, unlike the existing ML-based schemes of the literature, no PU statistical parameters are required. In order to discriminate between H0-class and elsewhere, we used a one-class classification approach that is inspired by the Isolation Forest algorithm. Extensive simulations are done in order to investigate the efficiency of such hybrid SS and the impact of the novelty detection model parameters on the detection performance. Indeed, these simulations corroborate the efficiency of the proposed one-class learning of the hybrid SS system.","PeriodicalId":6705,"journal":{"name":"2020 28th European Signal Processing Conference (EUSIPCO)","volume":"1 1","pages":"1683-1686"},"PeriodicalIF":0.0,"publicationDate":"2021-01-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"77083788","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-01-24DOI: 10.23919/Eusipco47968.2020.9287621
Mina Mounir, P. Karsmakers, T. Waterschoot
Detecting the onset of notes in music excerpts is a fundamental problem in many music signal processing tasks, including analysis, synthesis, and information retrieval. When addressing the note onset detection (NOD) problem using a data-driven methodology, a major challenge is the availability and quality of labeled datasets used for both model training/tuning and evaluation. As most of the available datasets are manually annotated, the amount of annotated music excerpts is limited and the annotation strategy and quality varies across data sets. To counter both problems, in this paper we propose to use semi-synthetic datasets where the music excerpts are mixes of isolated note recordings. The advantage resides in the annotations being automatically generated while mixing the notes, as isolated note onsets are straightforward to detect using a simple energy measure. A semi-synthetic dataset is used in this work for augmenting a real piano dataset when training a convolutional Neural Network (CNN) with three novel model training strategies. Training the CNN on a semi-synthetic dataset and retraining only the CNN classification layers on a real dataset results in higher average F1-score (F1) scores with lower variance.
{"title":"CNN-based Note Onset Detection using Synthetic Data Augmentation","authors":"Mina Mounir, P. Karsmakers, T. Waterschoot","doi":"10.23919/Eusipco47968.2020.9287621","DOIUrl":"https://doi.org/10.23919/Eusipco47968.2020.9287621","url":null,"abstract":"Detecting the onset of notes in music excerpts is a fundamental problem in many music signal processing tasks, including analysis, synthesis, and information retrieval. When addressing the note onset detection (NOD) problem using a data-driven methodology, a major challenge is the availability and quality of labeled datasets used for both model training/tuning and evaluation. As most of the available datasets are manually annotated, the amount of annotated music excerpts is limited and the annotation strategy and quality varies across data sets. To counter both problems, in this paper we propose to use semi-synthetic datasets where the music excerpts are mixes of isolated note recordings. The advantage resides in the annotations being automatically generated while mixing the notes, as isolated note onsets are straightforward to detect using a simple energy measure. A semi-synthetic dataset is used in this work for augmenting a real piano dataset when training a convolutional Neural Network (CNN) with three novel model training strategies. Training the CNN on a semi-synthetic dataset and retraining only the CNN classification layers on a real dataset results in higher average F1-score (F1) scores with lower variance.","PeriodicalId":6705,"journal":{"name":"2020 28th European Signal Processing Conference (EUSIPCO)","volume":"89 1","pages":"171-175"},"PeriodicalIF":0.0,"publicationDate":"2021-01-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"74046757","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：