Deep Neural Network Method Research Articles

Face Recognition of Indonesia’s Top Government Officials Using Deep Convolutional Neural Network

Facial recognition is a part of Computer Vision that is used to get facial coordinates from an image. Many algorithms have been developed to support Facial Detection such as Cascade Face Detection using Haar-Like features and AdaBoost to classify its Cascade and Convolutional Neural Network (CNN). Face recognition in this study uses the Deep Convolutional Neural Network (DCNN) method, and the output of this method is the measurement value of the face. In the model training process, Triplet Loss from Triplet Network Deep Metric Learning is used to get good face grouping results. The value of this face measurement will then be measured using the Euclidean distance calculation to determine the similarity of the input face from the dataset. This Research is using 6 images of Government officers in Indonesia to determine the accuracy of the model when there is a new picture of these officers inputted into the training machine. The result provides a 94% accuracy level with a variety of face positions and levels of brightness.

Quality Assessment Algorithm of X-Ray Images in Overall Girth Welds Based on Deep Neural Network

Defect recognition and assessment are essential technologies for the safety of pipelines. Recently, deep neural network methods have shown great progress in recognizing pipeline weld defects. However, the current methods cannot achieve online quality assessment for overall girth welds. To solve this problem, this paper proposes a novel quality assessment algorithm for defect recognition and assessment of X-ray girth weld images. First, a two-stage defect classifier based on residual neural network (ResNet) is proposed, which consists of a ResNet18 region classifier and a ResNet50 defect classifier. It can make up for the lack of a single classifier so that the performance and speed of defect recognition can be improved. Second, the feature-extraction calculator is used to extract and calculate the area of the weld defects based on the smallest bounding rectangle method. Third, we propose a quality evaluator based on an improved ensemble learning algorithm and long short-term memory (LSTM), so that the girth welds can be well assessed. Finally, an experiment is conducted using pipeline girth weld X-ray images. The analytic hierarchy process (AHP) method is used to comprehensively assess the level of overall girth-weld defects. The experimental results show that the proposed method is effective in recognizing and evaluating the girth-weld defects of X-ray images. And the accuracy of the quality evaluator can reach 92% for the quality assessment of overall girth welds, compared with manual quality decisions. It effectively achieves online quality assessment for girth welds and has higher evaluation accuracy.

Eigenvalue-Corrected Natural Gradient Based on a New Approximation

Using second-order optimization methods for training deep neural networks (DNNs) has attracted many researchers. A recently proposed method, Eigenvalue-corrected Kronecker Factorization (EKFAC), proposed an interpretation by viewing natural gradient update as a diagonal method and corrects the inaccurate re-scaling factor in the KFAC eigenbasis. What’s more, a new method to approximate the natural gradient called Trace-restricted Kronecker-factored Approximate Curvature (TKFAC) is also proposed, in which the Fisher information matrix (FIM) is approximated as a constant multiplied by the Kronecker product of two matrices and the traces can be kept equal before and after the approximation. In this work, we combine the ideas of these two methods and propose Trace-restricted Eigenvalue-corrected Kronecker Factorization (TEKFAC). The proposed method not only corrects the inexact re-scaling factor under the Kronecker-factored eigenbasis, but also considers the new approximation method and the effective damping technique adopted by TKFAC. We also discuss the differences and relationships among the related Kronecker-factored approximations. Empirically, our method outperforms SGD with momentum, Adam, EKFAC and TKFAC on several DNNs.

Novel biomolecular fingerprint using graph convolutional networks.

Molecular fingerprints are essential cheminformatics tools for machine learning with applications in drug discovery. Standard fingerprint software compute fixed-size feature vectors, and employ them as inputs to a deep neural network or other machine learning methods. Fixed-size fingerprint representation of molecules, however, requires extremely large vectors to encode all possible substructures. Limited accuracy and poor interpretation also occur due to the underlying neural network which emphasizes particular and exclusive aspects of the molecular structure. In this study, we develop a novel graph convolutional network (GCN) to predict the binding free energy of protein-ligand complexes. By adding a physics-based layer to the network architecture, the accuracy of the molecular fingerprint has been improved while potential overfitting has been avoided. In addition to standard information about the substuctures, our hyrbid physics-data model encodes atomic bonds features which captures structural features and enables further analysis. It has been shown that machine-optimized fingerprints, compared to fixed-sized fingerprints, can provide more accurate predictions, better performance, and more interpretable results. We show that the proposed GCN fingerprint outperforms the predictive performance of standard fingerprints on binding free energy of host-guest systems and PDBbind database.

Multi-Modality Deep Restoration of Extremely Compressed Face Videos.

Arguably the most common and salient object in daily video communications is the talking head, as encountered in social media, virtual classrooms, teleconferences, news broadcasting, talk shows, etc. When communication bandwidth is limited by network congestions or cost effectiveness, compression artifacts in talking head videos are inevitable. The resulting video quality degradation is highly visible and objectionable due to high acuity of human visual system to faces. To solve this problem, we develop a multi-modality deep convolutional neural network method for restoring face videos that are aggressively compressed. The main innovation is a new DCNN architecture that incorporates known priors of multiple modalities: the video-synchronized speech signal and semantic elements of the compression code stream, including motion vectors, code partition map and quantization parameters. These priors strongly correlate with the latent video and hence they are able to enhance the capability of deep learning to remove compression artifacts. Ample empirical evidences are presented to validate the superior performance of the proposed DCNN method on face videos over the existing state-of-the-art methods.

Dynamic Modeling and Model-Based Control with Neural Network-Based Compensation of a Five Degrees-of-Freedom Parallel Mechanism

In this paper, a spatial parallel mechanism with five degrees of freedom is studied in order to provide a promising dynamic model for the control design. According to the inverse kinematics of the mechanism, the dynamic model is derived by using the Lagrangian method, and the co-simulation using MSC ADAMS and MATLAB/Simulink is adopted to verify the established dynamic model. Then the pre-trained deep neural network (DNN) is introduced to predict the real-time state of the end-effector of the mechanism. Compared to the traditional Newton’s method, the DNN method reduces the cost of the forward kinematics calculation while ensuring prediction accuracy, which enables the dynamic compensation based on feedback signals. Furthermore, the computed torque control with DNN-based feedback compensation is implemented for the trajectory tracking of the mechanism. The simulations show that, in the most complicated case that involves friction and external disturbance, the proposed controller has better tracking performance. The results indicate the necessity of dynamic modeling in the design of control with high precision.

Multiple independent losses scheduling: A simple training method for deep neural networks

In recent years, various loss functions have been proposed to boost the performance of deep neural networks. Every loss function has its own specific theoretical motivation, and can easily learn its preference features of training data compared with other loss functions. Thus, combining multiple loss functions to capture more data features becomes an attractive idea for model performance improvement. In this paper, instead of using a single loss function or a linear weighted sum of multiple loss functions, we present the method named Multiple Independent Losses Scheduling (MILS), which allows multiple loss functions to independently participate in the training process according to their performance. Specifically, for all candidate loss functions, one loss function will be predefined as the primary loss function before training, and the other loss functions will play auxiliary roles for possible contributions to improve the model performance. In order to avoid auxiliary loss functions bringing a negative effect on the model performance in the training process, we developed a simple but effective performance-based scheduling algorithm to prevent auxiliary loss functions from dragging down the model performance. Extensive experiments using various deep architectures on various recognition benchmarks demonstrate our scheme is simple, robust, lightweight, and effective for typical classification tasks.

ResNet-SVM: Fusion based glioblastoma tumor segmentation and classification.

Computerized segmentation of brain tumor based on magnetic resonance imaging (MRI) data presents an important challenging act in computer vision. In image segmentation, numerous studies have explored the feasibility and advantages of employing deep neural network methods to automatically detect and segment brain tumors depicting on MRI. For training the deeper neural network, the procedure usually requires extensive computational power and it is also very time-consuming due to the complexity and the gradient diffusion difficulty. In order to address and help solve this challenge, we in this study present an automatic approach for Glioblastoma brain tumor segmentation based on deep Residual Learning Network (ResNet) to get over the gradient problem of deep Convolutional Neural Networks (CNNs). Using the extra layers added to a deep neural network, ResNet algorithm can effectively improve the accuracy and the performance, which is useful in solving complex problems with a much rapid training process. An additional method is then proposed to fully automatically classify different brain tumor categories (necrosis, edema, and enhancing regions). Results confirm that the proposed fusion method (ResNet-SVM) has an increased classification results of accuracy (AC = 89.36%), specificity (SP = 92.52%) and precision (PR = 90.12%) using 260 MRI data for the training and 112 data used for testing and validation of Glioblastoma tumor cases. Compared to the state-of-the art methods, the proposed scheme provides a higher performance by identifying Glioblastoma tumor type.

LA-ESN: A Novel Method for Time Series Classification

Time-series data is an appealing study topic in data mining and has a broad range of applications. Many approaches have been employed to handle time series classification (TSC) challenges with promising results, among which deep neural network methods have become mainstream. Echo State Networks (ESN) and Convolutional Neural Networks (CNN) are commonly utilized as deep neural network methods in TSC research. However, ESN and CNN can only extract local dependencies relations of time series, resulting in long-term temporal data dependence needing to be more challenging to capture. As a result, an encoder and decoder architecture named LA-ESN is proposed for TSC tasks. In LA-ESN, the encoder is composed of ESN, which is utilized to obtain the time series matrix representation. Meanwhile, the decoder consists of a one-dimensional CNN (1D CNN), a Long Short-Term Memory network (LSTM) and an Attention Mechanism (AM), which can extract local information and global dependencies from the representation. Finally, many comparative experimental studies were conducted on 128 univariate datasets from different domains, and three evaluation metrics including classification accuracy, mean error and mean rank were exploited to evaluate the performance. In comparison to other approaches, LA-ESN produced good results.

The classification of brain network for major depressive disorder patients based on deep graph convolutional neural network.

The early diagnosis of major depressive disorder (MDD) is very important for patients that suffer from severe and irreversible consequences of depression. It has been indicated that functional connectivity (FC) analysis based on functional magnetic resonance imaging (fMRI) data can provide valuable biomarkers for clinical diagnosis. However, previous studies mainly focus on brain disease classification in small sample sizes, which may lead to dramatic divergences in classification accuracy. This paper attempts to address this limitation by applying the deep graph convolutional neural network (DGCNN) method on a large multi-site MDD dataset. The resting-state fMRI data are acquired from 830 MDD patients and 771 normal controls (NC) shared by the REST-meta-MDD consortium. The DGCNN model trained with the binary network after thresholding, identified MDD patients from normal controls and achieved an accuracy of 72.1% with 10-fold cross-validation, which is 12.4%, 9.8%, and 7.6% higher than SVM, RF, and GCN, respectively. Moreover, the process of dataset reading and model training is faster. Therefore, it demonstrates the advantages of the DGCNN model with low time complexity and sound classification performance. Based on a large, multi-site dataset from MDD patients, the results expressed that DGCNN is not an extremely accurate method for MDD diagnosis. However, there is an improvement over previous methods with our goal of better understanding brain function and ultimately providing a biomarker or diagnostic capability for MDD diagnosis.

Bridge damage identification under the moving vehicle loads based on the method of physics-guided deep neural networks

Modeling uncertainty or modeling error has been widely recognized as one major challenge for damage identification of bridge structures by moving vehicles. This study proposed a physics-guided deep neural network approach to improve the accuracy of damage identification in vehicle-induced bridge vibration response, which aims to minimize the influence of the modeling uncertainty on damage identification. Firstly, the physical relationship between the damage to the bridge with the acceleration signal of the bridge, the physics-guided deep neural network, and its damage diagnosis method is introduced. The feasibility of the damage identification method based on physics-guided deep neural networks is theoretically explained. Secondly, a numerical case of damage identification of the simply supported bridge under moving vehicle loads is conducted to compare the accuracy of damage identification between the proposed method and the traditional method. Finally, a bridge field test is implemented to verify the applicability and effectiveness of the proposed method. The results illustrate that the method can be utilized to accurately identify the damage locations of simply supported bridges under a moving vehicle.

Deteksi Wajah Bermasker Menggunakan Deep Neural Network dan Tensorflow

The use of face masks is an important part of life in the midst of the Covid-19 pandemic which has been designated as a global pandemic, people are urged to cover their faces when in public areas to avoid the spread of the virus. The use of these face masks has raised serious questions regarding face detection systems. This study used Deep Neural Network and Tensorflow methods to detect faces both using masks and without masks. This study used two datasets, a face collection dataset without a mask and a face group using a mask. The result of this study was the detection of faces with masks trained on the dataset achieved 98% accuracy in training, and for testing got 100% accuracy on faces without masks, and 99.99% accuracy for face with mask.

Rock image classification using deep residual neural network with transfer learning

Rock image classification is a significant part of geological research. Compared with traditional image classification methods, rock image classification methods based on deep learning models have the great advantage in terms of automatic image features extraction. However, the rock classification accuracies of existing deep learning models are unsatisfied due to the weak feature extraction ability of the network model. In this study, a deep residual neural network (ResNet) model with the transfer learning method is proposed to establish the corresponding rock automatic classification model for seven kinds of rock images. ResNet34 introduces the residual structure to make it have an excellent effect in the field of image classification, which extracts high-quality rock image features and avoids information loss. The transfer learning method abstracts the deep features from the shallow features, and better express the rock texture features for classification in the case of fewer rock images. To improve the generalization of the model, a total of 3,82,536 rock images were generated for training via image slicing and data augmentation. The network parameters trained on the Texture Library dataset which contains 47 types of texture images and reflect the characteristics of rocks are used for transfer learning. This pre-trained weight is loaded when training the ResNet34 model with the rock dataset. Then the model parameters are fine-tuned to transfer the model to the rock classification problem. The experimental results show that the accuracy of the model without transfer learning reached 88.1%, while the model using transfer learning achieved an accuracy of 99.1%. Aiming at geological engineering field investigation, this paper studies the embedded deployment application of the rock classification network. The proposed rock classification network model is transplanted to an embedded platform. By designing a rock classification system, the off-line rock classification is realized, which provides a new solution for the rock classification problem in the geological survey. The deep residual neural network and transfer learning method used in this paper can automatically classify rock features without manually extracting. These methods reduce the influence of subjective factors and make the rock classification process more automatic and intelligent.

Analysis of Efficient and Fast Prediction Method for the Kinematics Solution of the Steel Bar Grinding Robot

Aiming at the robotization of the grinding process in the steel bar finishing process, the steel bar grinding robot can achieve the goal of fast, efficient, and accurate online grinding operation, a multi-layer forward propagating deep neural network (DNN) method is proposed to efficiently predict the kinematic solution of grinding robot. The process and kinematics model of the grinding robot are introduced. Based on the proposed method, simulations of the end position and orientation, and joint angle of the grinding robot are given. Three different methods, including SGD + tanh, Nadam + tanh, Nadam + ELU, are used to test the DNN calculation process results show that the method combining Nadam with ELU function has the fastest solution speed and higher accuracy can be obtained with the increase in iteration times. Finally, the Nadam optimizer is used to optimize the calculation results of the example. The optimization results show that this method accelerates the convergence rate of trajectory prediction error and improves the accuracy of trajectory prediction. Thus, the proposed method in this paper is an effective method to predict the kinematic solution when the grinding robot works online.

A Deep Learning-Based Method for Automatic Abnormal Data Detection: Case Study for Bridge Structural Health Monitoring

Ideally, the monitoring data collected by the Structural health monitoring (SHM) system should purely reflect the structure status. However, sensors deployed in the field can be very vulnerable to extreme conditions such as heavy rainfall, which results in large quantities of anomalous data and unavoidably leads to the inaccuracy of structural condition assessment and even false alarms. To automatically identify whether the collected data are abnormal or not, a novel deep learning-based data anomaly detection technique combining the time-frequency method and the Convolutional Neural Network (CNN) is proposed in this paper. First, the original time-series data of the SHM system were converted to the red green blue (RGB) images by using the wavelet scalograms. Subsequently, the GoogLeNet deep neural network is applied to construct a classification model by incorporating the generated 2D images. In order to evaluate the performance of the proposed technique, the SHM data (containing seven abnormal patterns) lasting for one month of a long-span cable-stayed bridge were utilized for experimental validation. The results indicate that compared with traditional deep neural network methods, the data anomaly identification accuracy can be improved by using the proposed technique. Different types of data anomaly patterns can be accurately identified, even in the case of small samples. The proposed technique exhibits good accuracy and can be integrated into advanced SHM systems with high fidelity and intelligence.

A systematic strategy of pallet identification and picking based on deep learning techniques

PurposeThis paper presents a comprehensive pallet-picking approach for forklift robots, comprising a pallet identification and localization algorithm (PILA) to detect and locate the pallet and a vehicle alignment algorithm (VAA) to align the vehicle fork arms with the targeted pallet.Design/methodology/approachOpposing vision-based methods or point cloud data strategies, we utilize a low-cost RGB-D camera, and thus PILA exploits both RGB and depth data to quickly and precisely recognize and localize the pallet. The developed method guarantees a high identification rate from RGB images and more precise 3D localization information than a depth camera. Additionally, a deep neural network (DNN) method is applied to detect and locate the pallet in the RGB images. Specifically, the point cloud data is correlated with the labeled region of interest (RoI) in the RGB images, and the pallet's front-face plane is extracted from the point cloud. Furthermore, PILA introduces a universal geometrical rule to identify the pallet's center as a “T-shape” without depending on specific pallet types. Finally, VAA is proposed to implement the vehicle approaching and pallet picking operations as a “proof-of-concept” to test PILA’s performance.FindingsExperimentally, the orientation angle and centric location of the two kinds of pallets are investigated without any artificial marking. The results show that the pallet could be located with a three-dimensional localization accuracy of 1 cm and an angle resolution of 0.4 degrees at a distance of 3 m with the vehicle control algorithm.Research limitations/implicationsPILA’s performance is limited by the current depth camera’s range (< = 3 m), and this is expected to be improved by using a better depth measurement device in the future.Originality/valueThe results demonstrate that the pallets can be located with an accuracy of 1cm along the x, y, and z directions and affording an angular resolution of 0.4 degrees at a distance of 3m in 700ms.

Pix2Pix and Deep Neural Network-Based Deep Learning Technology for Predicting Vortical Flow Fields and Aerodynamic Performance of Airfoils

Traditional computational fluid dynamics (CFD) methods are usually used to obtain information about the flow field over an airfoil by solving the Navier–Stokes equations for the mesh with boundary conditions. These methods are usually costly and time-consuming. In this study, the pix2pix method, which utilizes conditional generative adversarial networks (cGANs) for image-to-image translation, and a deep neural network (DNN) method were used to predict the airfoil flow field and aerodynamic performance for a wind turbine blade with various shapes, Reynolds numbers, and angles of attack. Pix2pix is a universal solution to the image-to-image translation problem that utilizes cGANs. It was successfully implemented to predict the airfoil flow field using fully implicit high-resolution scheme-based compressible CFD codes with genetic algorithms. The results showed that the vortical flow fields of the thick airfoils could be predicted well using the pix2pix method as a result of deep learning.

Evaluation of a Coating Process for SiO&lt;sub&gt;2&lt;/sub&gt;/TiO&lt;sub&gt;2&lt;/sub&gt; Composite Particles by Machine Learning Techniques

In this study, in order to optimize a fabrication process for SiO2/TiO2 composite particles and control their coating ratio (CTi), regression models for the coating process were constructed using various machine learning techniques. The composite particles with a core (SiO2)/shell (TiO2) structure were synthesized by mechanical stress under various fabrication conditions with respect to the supply volume of raw materials (V), addition ratio of TiO2 (rTi), operation time (t), rotor rotation speed (S), and temperature (T). Regression models were constructed by the least squares method (LSM), principal component regression (PCR), support vector regression (SVR), and the deep neural network (DNN) method. The accuracy of the constructed regression models was evaluated using the determination coefficients (R2) and the predictive performance was evaluated by comparing the prediction coefficients (Q2). From the perspective of the R2 and Q2 values, the DNN regression model was found to be the most suitable model for the present coating process. Moreover, the effects of the fabrication parameters on CTi were analyzed using the constructed DNN model. The results suggested that the t value was the dominant factor determining CTi of the composite particles, with the plot of CTi versus t displaying a clear maximum.

CLC-Net: Contextual and local collaborative network for lesion segmentation in diabetic retinopathy images

Diabetic retinopathy (DR) is the leading cause of blindness among people of working age. Fundus lesions are clinical signs of DR, and their recognition and delineation are important for early screening, grading, and monitoring of the disease. We propose in this work a fully automatic deep convolutional neural network method for simultaneous segmentation of four different types of DR-related fundus lesions. To exploit multi-scale image information, we propose a collaborative architecture that comprises a contextual branch and a local branch. An attention mechanism is designed to fuse feature maps from all decoding layers in order to effectively and fully combine informative features from the two branches. Moreover, an auxiliary classification task with a novel supervision scheme is introduced to reduce model overfitting and further improve the accuracy of lesion segmentation. Extensive experiments are conducted using three public fundus datasets, and our method produces a mean AUC value of 0.677, 0.629, and 0.581 on them respectively. The results demonstrate the advantages of the proposed method, outperforming alternative strategies and other state-of-the-art methods in the literature.

Structure genome based machine learning method for woven lattice structures

As a type of lightweight composite material, three-dimensional (3D) woven lattice structure (WLS) has been extensively applied in various fields. It is extremely significant to investigate its mechanical properties. Complex structure parameters make property prediction a great challenge. In this research, structure genome (SG) of 3D WLSs was defined with four SG characteristics, including the panel thickness, the meridional and zonal angles of fiber pillars and the mass fraction of the resin. A SG-based deep neural network method was conducted to predict the fundamental mechanical properties, including the flatwise compression strength and the out-of-plane elastic modulus. The effectiveness of the proposed method was verified by comparing the experimental data with the predictions. The results show that the prediction accuracy is more than 90% for most parts of the testing set. Through sensitivity analysis, the panel thickness and the mass fraction of the resin were identified as the most dominant parameter characteristics, which determine the overall performance. The method provides guidance for structure design of 3D WLSs.