- Conference Article
- 10.5339/qfarc.2016.ictpp1469
Classification of Bisyllabic Lexical Stress Patterns Using Deep Neural Networks
- Jan 01, 2016
- Mostafa Shahin + 1 more +1
Classification of Bisyllabic Lexical Stress Patterns Using Deep Neural Networks
Growing random forest on deep convolutional neural networks for scene categorization
Classification of Bisyllabic Lexical Stress Patterns Using Deep Neural Networks
Classification of Bisyllabic Lexical Stress Patterns Using Deep Neural Networks
Early detection of infestation by mustard aphid, vegetable thrips and two-spotted spider mite in bok choy with deep neural network (DNN) classification model using hyperspectral imaging data
Early detection of infestation by mustard aphid, vegetable thrips and two-spotted spider mite in bok choy with deep neural network (DNN) classification model using hyperspectral imaging data
Read moreAn Enhanced Diabetic Foot Ulcer Classification Approach Using GLCM and Deep Convolution Neural Network
An Enhanced Diabetic Foot Ulcer Classification Approach Using GLCM and Deep Convolution Neural Network
Mapping integrated crop-livestock systems in Brazil with planetscope time series and deep learning
Mapping integrated crop-livestock systems in Brazil with planetscope time series and deep learning
Sound Event Detection Using Derivative Features in Deep Neural Networks
We propose using derivative features for sound event detection based on deep neural networks. As input to the networks, we used log-mel-filterbank and its first and second derivative features for each frame of the audio signal. Two deep neural networks were used to evaluate the effectiveness of these derivative features. Specifically, a convolutional recurrent neural network (CRNN) was constructed by combining a convolutional neural network and a recurrent neural networks (RNN) followed by a feed-forward neural network (FNN) acting as a classification layer. In addition, a mean-teacher model based on an attention CRNN was used. Both models had an average pooling layer at the output so that weakly labeled and unlabeled audio data may be used during model training. Under the various training conditions, depending on the neural network architecture and training set, the use of derivative features resulted in a consistent performance improvement by using the derivative features. Experiments on audio data from the Detection and Classification of Acoustic Scenes and Events 2018 and 2019 challenges indicated that a maximum relative improvement of 16.9% was obtained in terms of the F-score.
Read moreDiverse Feature Visualizations Reveal Invariances in Early Layers of Deep Neural Networks
Visualizing features in deep neural networks (DNNs) can help understanding their computations. Many previous studies aimed to visualize the selectivity of individual units by finding meaningful images that maximize their activation. However, comparably little attention has been paid to visualizing to what image transformations units in DNNs are invariant. Here we propose a method to discover invariances in the responses of hidden layer units of deep neural networks. Our approach is based on simultaneously searching for a batch of images that strongly activate a unit while at the same time being as distinct from each other as possible. We find that even early convolutional layers in VGG-19 exhibit various forms of response invariance: near-perfect phase invariance in some units and invariance to local diffeomorphic transformations in others. At the same time, we uncover representational differences with ResNet-50 in its corresponding layers. We conclude that invariance transformations are a major computational component learned by DNNs and we provide a systematic method to study them.
Read moreContact Failure Diagnosis for GIS Plug-In Connector by Magnetic Field Measurements and Deep Neural Network Classifiers
This article presents a contact fault diagnosis method of gas-insulated switchgear (GIS) plug-in connector via magnetic field measurement, magnetic field visualization, and deep neural network (DNN) classifiers. First, the surrounding magnetic field of GIS plug-in connector with normal contact (NC) condition and with artificially designed contact failures was measured by the Hall sensor array. Then, the measured magnetic field was gathered with an original matrix of 16 <inline-formula> <tex-math notation="LaTeX">$\times$</tex-math> </inline-formula> 16 dimensions. The <inline-formula> <tex-math notation="LaTeX">$\vert$</tex-math> </inline-formula>original matrix was then visualized by the max–min normalization and correlation matrix. Database containing 11 000 magnetic field images was labeled and segmented as training, validation, and test datasets. Furthermore, high-dimensional features of input magnetic field images were extracted by different DNN filters, including convolutional neural network (CNN), simple recurrent neural network (Sim-RNN), and long short-term memory (LSTM) network. Then, extracted high-dimensional features were fed into a fully connected (Fc) neural network with SoftMax classifiers to identify different contact faults. Finally, the performance of different DNN-based classifiers is compared by the fault classification merits, <inline-formula> <tex-math notation="LaTeX">$t$</tex-math> </inline-formula>-distributed stochastic neighbor embedding (<inline-formula> <tex-math notation="LaTeX">$t$</tex-math> </inline-formula>-SNE) feature clustering, and confusion matrixes. Results show that the DNN-based model could achieve contact fault classification task with an accuracy of 97.7% and <inline-formula> <tex-math notation="LaTeX">$F_{1\_\mathrm{score}}$</tex-math> </inline-formula> of 0.985. Therefore, the proposed method is useful for designing a high-performance contact status monitoring system of GIS equipment, thus improving its operation safety.
Read moreDevelopment of a Radiomic Signature from Pre-Treatment FDG-PET for Recurrence Prediction in Cervical Cancer.
Development of a Radiomic Signature from Pre-Treatment FDG-PET for Recurrence Prediction in Cervical Cancer.
A Study of the Computation Amount and Computation Time of Classical Deep Neural Networks
Deep convolutional neural networks have made great progress in a variety of computer vision tasks. With the gradual improvement of its performance, the layers of neural networks become deeper, and the training-validation time and computational complexity increase dramatically. How to find the relationship between characteristics of deep neural networks and their training-validation time is of great significance for accelerating convolutional neural networks. This paper analyzes the computation of several classical deep convolutional neural networks (DNNs) proposed in the field of image recognition, and compares the computation with the total time of actual training and verification. It is found that the computation of a deep neural network is not linearly related to running time of training and verification.
Read moreDeep Learning Based Emotion Classification Using Mel Frequency Magnitude Coefficient
The popularity of emotion recognition using speech signals increases more and more because of its vast number of applications in the practical field. Emotion recognition using the speech signal is a very complicated and challenging task. This plays an essential role in enhancing human-computer interaction (HCI). Many authors used different methods to improve the accuracy of speech emotion recognition (SER). Proper selection of features and suitable machine and deep learning model design can improve the recognition rate. In this work, we used a modified version of the mel frequency cepstral coefficient (MFCC) feature named the mel frequency magnitude coefficient (MFMC) with convolutional neural network (CNN) and deep neural network (DNN) classifiers to enhance the SER. We used MFMC and MFCC features as input to CNN and DNN classifiers and evaluated the accuracy of SER. We made two observations from our experiment. First, the performance of the MFMC feature in SER is better than the MFCC feature for both classifiers. Second, the proposed DNN classifier achieved better accuracy than the CNN classifier for both features (MFMC and MFCC). The MFMC feature with the DNN classifier achieved an accuracy of 76.72%, 84.72%, 77.88%, and 100% for the RA VDESS, EMODB, SA VEE, and TESS datasets, respectively. Similarly, the CNN classifier with the MFMC feature achieved an accuracy of 72.9%, 82.41 %, 74.5%, and 100% for the same datasets. Our proposed work was compared with the state-of-the-art models, and we found that our model performed better than others.
Read moreDNN Classifier and Decision Tree-Based Novel Methodology for Analyzing Road Accidents
Nation's social and economic improvement is highly based on the citizen's health profile. Road safety can be considered as one of the most crucial and vital aspects of every country. The trauma of road accidents is, it mainly affects developing countries like Sri Lanka, rather than the developed countries. In Sri Lanka, 2.8% of the deaths are occurred by road accidents which indicate the number of 3554 deaths in last year and over 180 US $ million annually have to spend on road accidents. Furthermore, to reduce the danger of road accidents and to improve road safety, an analysis of road accidents based on machine learning was proposed. The primary aim of this study is to develop a mechanism to predict the driver fault at the accident and identify the accident hot spot in the area, which can be effectively used for taking countermeasures to improve road safety in Sri Lanka. Deep Neural Networks (DNN) classifier, Decision tree, Linear classifier in deep learning have been used to accurately predict and uncover road accident patterns within the data set. Results revealed that the Decision Tree and the DNN classifier are more accurate and comparable than other machine learning techniques used and the first 25 km of the road is more accident-prone rather the other road segments in the road. The research will provide the information needed to guide the relevant decision- makers in adopting suitable measures to prevent and to reduce the accident rate.
Read moreLearning with noisy labels using collaborative sample selection and contrastive semi-supervised learning
Learning with noisy labels using collaborative sample selection and contrastive semi-supervised learning
Intelligent recognition of ancient murals based on deep learning models
This study focuses on the digital protection needs of ancient Chinese mural art and innovatively constructs an intelligent recognition system based on deep transfer learning and multi-modal feature fusion. In response to the challenges of scarce and uneven quality of mural image data, the system adopts a two-stage progressive training strategy. Firstly, a deep convolutional neural network is pre-trained on a large general image dataset to obtain the basic visual feature representation and then fine-tuned and optimized on a professional mural dataset through a domain adaptive method. The feature extraction stage innovatively integrates the high-level semantic features of deep neural networks with the low-level artistic features of traditional image processing. Among them, the deep features are extracted through an improved residual network architecture, while the artistic features blend the color distribution features based on the HSV color space with the improved rotational invariance LBP texture features. The system innovatively designs a feature fusion module based on the attention mechanism, achieving dynamic optimization and combination of multi-source features through learnable feature weights. Experimental verification shows that, compared with mainstream deep network models, this system has achieved a significant improvement in the accuracy of feature recognition in the era of murals. At the same time, it has demonstrated excellent generalization performance in cross-dataset tests, providing reliable technical support and new methodological guidance for the intelligent research of cultural heritage.
Read moreSpeech enhancement from fused features based on deep neural network and gated recurrent unit network
Speech is easily interfered by external environment in reality, which results in the loss of important features. Deep learning has become a popular speech enhancement method because of its superior potential in solving nonlinear mapping problems for complex features. However, the deficiency of traditional deep learning methods is the weak learning capability of important information from previous time steps and long-term event dependencies between the time-series data. To overcome this problem, we propose a novel speech enhancement method based on the fused features of deep neural networks (DNNs) and gated recurrent unit (GRU). The proposed method uses GRU to reduce the number of parameters of DNNs and acquire the context information of the speech, which improves the enhanced speech quality and intelligibility. Firstly, DNN with multiple hidden layers is used to learn the mapping relationship between the logarithmic power spectrum (LPS) features of noisy speech and clean speech. Secondly, the LPS feature of the deep neural network is fused with the noisy speech as the input of GRU network to compensate the missing context information. Finally, GRU network is performed to learn the mapping relationship between LPS features and log power spectrum features of clean speech spectrum. The proposed model is experimentally compared with traditional speech enhancement models, including DNN, CNN, LSTM and GRU. Experimental results demonstrate that the PESQ, SSNR and STOI of the proposed algorithm are improved by 30.72%, 39.84% and 5.53%, respectively, compared with the noise signal under the condition of matched noise. Under the condition of unmatched noise, the PESQ and STOI of the algorithm are improved by 23.8% and 37.36%, respectively. The advantage of the proposed method is that it uses the key information of features to suppress noise in both matched and unmatched noise cases and the proposed method outperforms other common methods in speech enhancement.
Read moreA Comparative Study of Deep Learning Models With Handcraft Features and Non-Handcraft Features for Automatic Plant Species Identification
The classification of plants is one of the most important aims for botanists since plants have a significant part in the natural life cycle. In this work, a leaf-based automatic plant classification framework is investigated. The aim is to compare two different deep learning approaches named Deep Neural Network (DNN) and deep Convolutional Neural Network (CNN). In the case of deep neural network, hybrid shapes and texture features are utilized as hand-crafted features while in the case of the convolution non-handcraft, features are applied for classification. The offered frameworks are evaluated with a public leaf database. From the simulation results, it is confirmed that the deep CNN-based deep learning framework demonstrates superior classification performance than the handcraft feature based approach.
Read more