Showing 6 results for Convolutional Neural Network
B. Nasersharif, N. Naderi,
Volume 17, Issue 2 (6-2021)
Abstract
Convolutional Neural Networks (CNNs) have been shown their performance in speech recognition systems for extracting features, and also acoustic modeling. In addition, CNNs have been used for robust speech recognition and competitive results have been reported. Convolutive Bottleneck Network (CBN) is a kind of CNNs which has a bottleneck layer among its fully connected layers. The bottleneck features extracted by CBNs contain discriminative and rich context information. In this paper, we discuss these bottleneck features from an information theory viewpoint and use them as robust features for noisy speech recognition. In the proposed method, CBN inputs are the noisy logarithm of Mel filter bank energies (LMFBs) in a number of neighbor frames and its outputs are corresponding phone labels. In such a system, we showed that the mutual information between the bottleneck layer and labels are higher than the mutual information between noisy input features and labels. Thus, the bottleneck features are a denoised compressed form of input features which are more representative than input features for discriminating phone classes. Experimental results on the Aurora2 database show that bottleneck features extracted by CBN outperform some conventional speech features and also robust features extracted by CNN.
P. Ramezanpour, M. Aghababaie, M. R. Mosavi, D. M. de Andrés,
Volume 18, Issue 2 (6-2022)
Abstract
Through beamforming, the desired signal is estimated by calculating the weighted sum of the input signals of an array of antenna elements. In the classical beamforming methods, computing the optimal weight vector requires prior knowledge on the direction of arrival (DoA) of the desired signal sources. However, in practice, the DoA of the signal of interest is unknown. In this paper, we introduce two different deep-neural-network-based beamformers which can estimate the signal of interest while suppressing noise and interferences in two/three stages when the DoAs are unknown. Employing deep neural networks (DNNs) such as convolutional neural networks (CNNs) and bidirectional long short-term memory (bi-LSTM) networks enables the proposed method to have better performance than existing methods. In most cases, the output signal to interference and noise ratio (SINR) of the proposed beamformer is more than 10dB higher than the output SINR of the classical beamformers.
A. Ataee, S. J. Kazemitabar,
Volume 19, Issue 1 (3-2023)
Abstract
We propose a real-time Yolov5 based deep convolutional neural network for detecting ships in the video. We begin with two famous publicly available SeaShip datasets each having around 9,000 images. We then supplement that with our self-collected dataset containing another thirteen thousand images. These images were labeled in six different classes, including passenger ships, military ships, cargo ships, container ships, fishing boats, and crane ships. The results confirm that Yolov5s can classify the ship's position in real-time from 135 frames per second videos with 99 % precision.
O. Mahmoudi Mehr, M. R. Mohammadi, M. Soryani,
Volume 19, Issue 3 (9-2023)
Abstract
Speckle noise is an inherent artifact appearing in medical images that significantly lowers the quality and accuracy of diagnosis and treatment. Therefore, speckle reduction is considered as an essential step before processing and analyzing the ultrasound images. In this paper, we propose an ultrasound speckle reduction method based on speckle noise model estimation using a deep learning architecture called “speckle noise-based inception convolutional denoising neural network" (SNICDNN). Regarding the complicated nature of speckle noise, an inception module is added to the first layer to boost the power of feature extraction. Reconstruction of the despeckled image is performed by introducing a mathematical method based on solving a quadratic equation and applying an image-based inception convolutional denoising autoencoder (IICDAE). The results of various quantitative and qualitative evaluations on real ultrasound images demonstrate that SNICDNN outperforms the state-of-the-art methods for ultrasound despeckling. SNICDNN achieves 0.4579 dB and 0.0100 additional gains on average for PSNR and SSIM, respectively, compared to other methods. Denoising ultrasound based on its noise model estimation is not only a novel approach in comparison to traditional denoising autoencoder models but also due to the fact that it uses mathematical solutions to recover denoised images, SNICDNN shows a greater power in ultrasound despeckling.
V. Esmaeili, M. Mohassel Feghhi,
Volume 19, Issue 3 (9-2023)
Abstract
The coronavirus disease or COVID-19, as a global disease, is an unprecedented health care crisis due to increasing mortality and its high rate of infection. Patients usually show significant complications in the respiratory system. This disease is caused by SARS-CoV-2. Decreasing the time of diagnosis is essential for reducing deaths and low spreading of the virus. Also, using the optimal tool in the pediatric setting and Intensive care unit (ICU) is required. Therefore, using lung ultrasound is recommended. It does not have any radiation and it has a lower cost. However, it makes noisy and low-quality data. In this paper, we propose a novel approach called Uniform Local Binary Pattern on Five intersecting Planes and convolutional neural Network (ULBPFP-Net) that overcomes the said limitation. We extract worthwhile features from five planes for feeding a network. Our experiments confirm the success of the ULBPFP-Net in COVID-19 diagnosis compared to the previous approaches.
Robinson Jimenez-Moreno, Anny Astrid Espitia Cubillos, Esperanza Rodríguez Carmona,
Volume 20, Issue 0 (12-2024)
Abstract
This document presents the design of a virtual robotic system for the supervision of physical training exercises, to be carried out in a closed environment, which only requires a computer equipment with a web camera. To do this, deep learning algorithms such as convolutional networks and short- and long-term memory networks are used to recognize voice commands and the user's video actions. A predefined dialogue template is used to guide a user's training cycle based on the execution of the exercises: push-ups, abdominal, jump or squat. The contribution of the work focuses on the integration of deep learning techniques to design and personalize virtual robotic assistants for everyday task. The results show a high level of accuracy by the virtual robot both in understanding the audio and in predicting the exercise to be performed, with a final accuracy value of 97.75% and 100%, respectively.
