Dezhi Wang

Dr. Dezhi Wang


Research Focus & Metrics

Academic Career Length

9Years (2018-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
18 September 2018
Verified
Book Chapter
Closed access
PDF file
KXKele Xu
Kele Xu
DFDawei Feng
Dawei Feng
HMHaibo Mi
Haibo Mi
BZBoqing Zhu
Boqing Zhu
DWDezhi Wang
Dezhi Wang
SLShuwen Liu
Shuwen Liu

12 March 2019 in Applied Sciences
Verified
Journal Article
Open access
PDF file
LZLilun Zhang
Lilun Zhang
DWDezhi Wang
Dezhi Wang
CBChangchun Bao
Changchun Bao
YWYongxian Wang
Yongxian Wang
KXKele Xu
Kele Xu
Whale vocal calls contain valuable information and abundant characteristics that are important for classification of whale sub-populations and related biological research. In this study, an effective data-driven approach based on pre-trained Convolutional Neural Networks (CNN) using multi-scale waveforms and time-frequency feature representations is developed in order to perform the classification of whale calls from a large open-source dataset recorded by sensors carried by whales. Specifically, the classification is carried out through a transfer learning approach by using pre-trained state-of-the-art CNN models in the field of computer vision. 1D raw waveforms and 2D log-mel features of the whale-call data are respectively used as the input of CNN models. For raw waveform input, windows are applied to capture multiple sketches of a whale-call clip at different time scales and stack the features from different sketches for classification. When using the log-mel features, the delta and delta-delta features are also calculated to produce a 3-channel feature representation for analysis. In the training, a 4-fold cross-validation technique is employed to reduce the overfitting effect, while the Mix-up technique is also applied to implement data augmentation in order to further improve the system performance. The results show that the proposed method can improve the accuracies by more than 20% in percentage for the classification into 16 whale pods compared with the baseline method using groups of 2D shape descriptors of spectrograms and the Fisher discriminant scores on the same dataset. Moreover, it is shown that classifications based on log-mel features have higher accuracies than those based directly on raw waveforms. The phylogeny graph is also produced to significantly illustrate the relationships among the whale sub-populations.

19 September 2018
Verified
Book Chapter
Closed access
PDF file
BZBoqing Zhu
Boqing Zhu
KXKele Xu
Kele Xu
DWDezhi Wang
Dezhi Wang
LZLilun Zhang
Lilun Zhang
BLBo Li
Bo Li
YPYuxing Peng
Yuxing Peng

1 February 2019 in Journal of Physics: Conference Series
Verified
Conference Paper
Open access
PDF file
DWDezhi Wang
Dezhi Wang
LZLilun Zhang
Lilun Zhang
KXKele Xu
Kele Xu
YWYongxian Wang
Yongxian Wang
Motivated by the state-of-the-art performance of Dense Convolutional Networks (DenseNet) on highly competitive object recognition benchmark tasks (CIFAR-10, CIFAR-100, SVHN and ImageNet), this work presents improvements to the Acoustic Scene Classification (ASC) task of the IEEE AASP Challenge on Detection and Classification of Acoustic Scenes and Events (DCASE 2017) based on an optimized DenseNet model. Multi-channel Convolutional Neural Network (CNN) is also explored to extract features from different audio channels in an end-to-end manner. In the experiments, the proposed model is compared with the challenge baseline model of DCASE 2017 and several other state-of-the-art CNN architectures subject to the classification accuracy on the same open-source DCASE datasets. The results show that the proposed DenseNet-based architecture can achieve a superior performance in classification accuracy but with a lower model complexity in contrast with other models.

10 January 2018 in DEStech Transactions on Computer Science and Engineering
Verified
Journal Article
Closed access
XGXianPeng Guo
XianPeng Guo
LZLilun Zhang
Lilun Zhang
DWDezhi Wang
Dezhi Wang
ZLZengquan Lu
Zengquan Lu
YWYongxian Wang
Yongxian Wang
JZJun Zhang
Jun Zhang
In this paper, a Kriging surrogate model is introduced to simplify the computation of underwater acoustic propagation and quantify the related uncertainty. By using a Kriging model to replace the single frequency and multi-frequency FOR3D acoustic model, the computational speed is increased by dozens of times while ensuring that the maximum mean square error is confined within 5%. In this way, the balance between efficiency and accuracy of the underwater acoustic propagation is achieved. On the basis of this Kriging surrogate model, the Monte Carlo simulation is also used to analyze the uncertainty of acoustic propagation loss, in terms of the posterior probability density distribution. Finally, the confidence intervals corresponding to different confidence levels are calculated in order to quantify the uncertainty of transmission loss.

1 September 2018 in The Journal of the Acoustical Society of America
Verified
Journal Article
Closed access
PDF file
DWDezhi Wang
Dezhi Wang
LZLilun Zhang
Lilun Zhang
CBChangchun Bao
Changchun Bao
KXKele Xu
Kele Xu
BZBoqing Zhu
Boqing Zhu
ZLZengquan Lu
Zengquan Lu
In oceanic remote sensing, large discrete linear hydrophone arrays are usually utilized to enhance the signal-to-noise ratio (SNR) by means of beamforming that attenuates noise coming from the directions outside the target direction. Although the widely adopted delay-and-sum (DS) and filter-and-sum (FS) beamforming techniques can improve the performance of linear hydrophone arrays in some scenarios, both DS and FS methods have limited adaptability to the changing oceanic environments especially with spatially correlated noise. Moreover, these beamforming techniques aims to optimize the SNR, which is not completely consistent with some objectives such as target recognition. In this work, a neural network adaptive beamforming framework for an uniformly-spaced linear hydrophone array is proposed to make use of the large-scale array signals and address the above issues. In particular, an architecture consists of temporal convolutional neural networks is designed to predict the beamforming filter coefficients in time domain by taking the raw multi-channel waveforms of sound pressure as the input. The filter prediction networks are also jointly trained with a convolutional neural network (CNN) based classification model with the purpose of increasing the target recognition accuracy. The proposed approach is validated by experiments carried out in the south coastal waters of China.