Search

In this research, we delve into the unexplored potential of Long Short-term Memory (LSTM) and Gated Recurrent Unit (GRU) networks in the field of Computer Vision (CV). With the trend of Deep Learning methods shifting towards complex Neural Network architectures, researchers have acknowledged the efficacy of Dense and Convolutional Neural Networks (CNNs) in CV tasks, while Recurrent Neural Networks (RNNs) have been proven to be effective in Natural Language Processing (NLP) problems. Despite some recent developments in blending RNN with CNN for NLP, there have been very few attempts in applying RNN-based approaches in CV. To address this gap, we take a unique in order to develop LSTM and GRU models by tuning the number of layers and number of neurons in each layer. The models are then applied to multiple benchmark datasets (Fashion-MNIST, MNIST, EMNIST-Letters, EMNIST-Digits, Kuzushiji-MNIST, and Kuzushiji-49). Our experimental results indicate that LSTM and GRU networks significantly outperforms corresponding CNN and Dense networks with optimal number of parameters. Our findings suggest that RNN (LSTM and GRU) can be valuable alternatives to CNN and Dense networks for image classification problems in CV. This study presents a new direction for further research in the field and highlights the potential of LSTM and GRU networks in CV tasks.

Direct impingement (DI) erosion testing of steel and cobalt-based alloys is conducted with particles entrained in air and water. The tests are performed using 75- and 300-μm silica particles with three particle velocities and five impact angles. The mass loss results are used to develop an empirical equation that contains an explicit exponential term for the particle size effect. The erosion equation obtained from the gas jet experiments is implemented into a CFD solver with Lagrangian particle tracking for the liquid-submerged geometry. The CFD-based erosion predictions agree fairly with the experimental measurement for both particles. Strong effect has been observed for the exponential correction term in the erosion equation for liquid-solid flow cases; particularly with small particles at low impact velocities.

The application of cross-linking resin is an effective method for improving and controlling dimensional stability, such as the shrinkage of viscose single jersey knits. However, such treatment often leads to a significant deterioration in the bursting strength of treated fabrics. In this regard, resin treatment using a softening agent can be an additional potential solution for retaining the bursting strength of treated fabrics. Resin treatment is one kind of chemical finishing process that inhibits cellulosic textile fibre swelling during wetting, provides fibre resistance to deformation and prevents shrinkage. The key objective of this study was to model the effect of resin-finishing process variables for predicting the shrinkage control and bursting strength of viscose single jersey knitted fabrics. The MATLAB (Version 8.2.0.701) fuzzy expert system was used to model the optimum resin and softener concentrations, as well as the best curing time for the prediction of maximum shrinkage control with a minimum loss in fabric bursting strength. The optimal process variables were found to be a resin concentration of 75 g/l, a softener concentration of 45 g/l and a curing time of 225 seconds. The fuzzy expert model developed in this study was validated using experimental data. It was found that the model has the ability and accuracy to predict fabric shrinkage and bursting strength effectively in the non-linear field.

Sentiment analysis is a fundamental part of Natural Language Processing. There are numerous works on this topic in English and other languages. However, it is still a comparatively new practice in Bangla. The absence of a suitable Bangla corpus is the primary obstacle for sentiment analysis tasks in Bangla. Nonetheless, Long Short-Term Memory (LSTM) is a common technique for resolving sentiments from a dataset containing a large amount of text data. However, Gated Recurrent Unit (GRU) is very efficient for datasets with a low amount of text data. In this manuscript, we present a five-layered GRU neural network model, each layer comprising of 48 neurons, applied the model on an existing Bangla corpus. We have implemented the ten-fold cross-validation approach and repeated the same processes three times. Each time, we have considered the averages of the ten validation accuracy and losses and compared the results with the state-of-the-art published outcome (77.85% highest accuracy) for Bidirectional LSTM (BLSTM). The highest accuracy for our model is 78.41%, while the lowest accuracy is 76.34%.

Skin cancer is one of the worst diseases noticed in humankind. It beholds some types, which even experts find challenging to categorize. In recent times, neural network-based automated systems have been entitled to perform this difficult task for their amazing ability of pattern recognition. However, the challenge remains due to the requirement for high-quality images and thus the necessity of highly configured resources. In this research manuscript, the authors have addressed these issues. They pushed the boundary of neural networks by utilizing a low-resolution (80 × 80, 64 × 64, and 32 × 32 pixels), highly imbalanced, grayscale HAM10000 skin cancer dataset into several pre-trained network architectures (VGG16, DenseNet169, DenseNet161, and ResNet50) that have been successfully used for a similar purpose with a high-resolution, augmented RGB HAM10000 skin cancer image dataset. The image resolution of the original HAM10000 dataset is 800 × 600 pixels. The highest achieved performance for 80 × 80, 64 × 64, and 32 × 32 pixel images were 80.46%, 78.56%, and 74.15%, respectively. All of these results were accomplished from the ImageNet pre-trained VGG16 model. The second-best model in terms of transfer learning was DenseNet169. The performances demonstrate that even within these severe circumstances, neural network-based transfer learning holds promising possibilities.

This paper introduces a dataset of leaf vein images from four different species (two from Monocotyledons and two from Dicotyledons). Then multiple instances of the dataset containing 64 × 64, 32 × 32, 16 × 16, 8 × 8, and 4 × 4 pixel single channel center-focused images were created, and a new Residual Neural Network (ResNet) based model has applied to each of the instances for cotyledon (seed leaf or embryonic leaf) group identification. Additionally, the same procedures were followed for plant species classification. The aim was to make use of only the vein patterns for the task. The results show that despite the difficulties of recognizing patterns from small-scale images, it is likely to efficiently categorize seed types and plant species by properly deploying residual blocks in neural network models. Also, the applied ResNet was compared against ResNet-152 V2 on the 64 × 64 pixel imageset in terms of species classification.

In this research work, Galerkin and collocation methods have been introduced for approximating the solution of FIE of 2nd kind using LH (product of Laguerre and Hermite) polynomials which are considered as basis functions. Also, a comparison has been done between the solutions of Galerkin and collocation method with the exact solution. Both of these methods show the outcome in terms of the approximate polynomial which is a linear combination of basis functions. Results reveal that performance of collocation method is better than Galerkin method. Moreover, five different polynomials such as Legendre, Laguerre, Hermite, Chebyshev 1st kind and Bernstein are also considered as a basis functions. And it is found that all these approximate solutions converge to same polynomial solution and then a comparison has been made with the exact solution. In addition, five different set of collocation points are also being considered and then the approximate results are compared with the exact analytical solution. It is observed that collocation method performed well compared to Galerkin method. GANIT J. Bangladesh Math. Soc.Vol. 38 (2018) 11-25

Due to the ability to model various complex phenomena where classical calculus failed, fractional calculus is getting enormous attention recently. There are several approaches available for numerical approximations of various types of fractional differential equations. For fractional diffusion equations spectral collocation is one of the efficient and most popular ap-proximation techniques. In this research, we introduce spectral collocation method based on Lagrange’s basis polynomials for numerical approximations of two-dimensional (2D) space fractional diffusion equations where spatial fractional derivative is described in Riemann-Liouville sense. We consider four different types of nodes to generate Lagrange’s basis polynomials and as collocation points in the proposed spectral collocation technique. Spectral collocation method converts the diffusion equation into a system of ordinary differential equations (ODE) for time variable and we use 4th order Runge-Kutta method to solve the resulting system of ODE. Two examples are considered to verify the efficiency of different types of nodes in the proposed method. We compare approximated solution with exact solution and find that Lagrange’s spectral collocation method gives very high accuracy approximation. Among the four types of nodes, nodes from Jacobi polynomial give highest accuracy and nodes from Chebyshev polynomials of 1st kind give lowest accuracy in the proposed method.

Isocitrate dehydrogenase (IDH) isoform 1 (IDH1) and isoform 2 (IDH2) genes code for proteins that play important roles in the decarboxylation of isocitrate to α-ketoglutarate. Somatic mutations in the IDH1 and IDH2 genes are prognostically significant in patients with glioma and acute myeloid leukemia. However, very little is known about the germline polymorphisms in these two genes. In this study, we aim to comprehensively characterize the single nucleotide polymorphisms (SNPs) located in the IDH1 and IDH2 genes using in silico tools. For this purpose, the data for the SNPs were retrieved from the dbSNP database. The coding non-synonymous SNPs were analyzed for their structural and functional significance using a set of bioinformatics tools including SIFT, PolyPhen-2, PROVEAN, PhD-SNP, PMut, I-Mutant 3.0, MUpro and mutation3D. Structural comparison between normal and mutated residues for the disease-associated functional SNPs was performed using TM-align and SWIS MODEL. The non-coding SNPs located at the 3′-UTR were analyzed using PolymiRTS for assessing the impact of these SNPs in the regulation of binding sites for miRNAs. Moreover, the non-coding SNPs were further analyzed using HaploReg and RegulomeDB databases for assessing SNPs in the context of their chromatin state, regulation of protein binding and associated regulatory potential. A total of 26 SNPs from the IDH1 and 17 SNPs from the IDH2 gene were predicted to be functionally deleterious, damaging and disease associated. Cluster analysis showed that the IDH1 and IDH2 genes have SNPs (R108L, D273G, D279G at the IDH1 and V305M, R377H at the IDH2) clustered in the active site of these enzymes, further potentiating their significance. In these two clusters, the hotspot positions for cancer-specific somatic mutations are also located (R132H for IDH1 and R172H for IDH2). Moreover, SNPs affecting metal binding were also observed at positions 220, 273, and 306 of IDH1 and positions 261 and 356 of IDH2. Analysis of the non-coding SNPs predicted a number of SNPs to be located in the miRNA binding sites and at regulatory loci. In summary, this is the first comprehensive in silico study for the characterization of the germline polymorphisms in these genes and these data should further be validated in clinical studies.

Nowadays, SMS based communication is common in almost every sector. Private and government organizations deliver and receive millions of SMSs every year. This huge amount of SMSs can be used later in several useful ways. However, the storage space required for storing these huge amount SMSs creates so much complexity for the SMS service providing organizations that they cannot save the details of every SMS. Therefore, in this paper, we introduce a template-oriented SMS storage system based on a graph-theoretic approach, which reduces storage space requirements significantly. We achieve the storage space reduction through considering the fixed parts of SMSs as graph nodes and variable parts as graph edges. We perform experimentation on real SMS data collected from different telecom operators of Bangladesh. Our experimentation confirms that our proposed system substantially reduces the storage space requirement. To the best of our knowledge, we are the first to propose such a space-saving system for storing SMSs.

Social media has become important for sharing contents and people around the globe are expressing their feelings over social media. Analyzing these social interactions can be very much helpful to figure out the reasons behind people's feelings. In this work, we applied Long Short-term Memory (LSTM) Neural Networks for analyzing negative sentences in Bangla. We extracted negative Bangla tweets from twitter and also collected more negative Bangla sentences from people using google form. In our work, we have shown how hyper-parameters tuning can be useful for analyzing a small dataset and then applied to-fold cross-validation. Finally, we showed a relative comparison between our research and other previous works. This result will help psychologists to find out the reasons behind these social media interactions and prevent undesirable happenings resulted from negativity.

Human emotions like depression are inner sentiments of human beings which expose actual behaviors of a person. Analyzing and determining these type of emotions from people’s social activities in virtual world can be very helpful to understand their behaviors. Existing approaches may be useful for analyzing common sentiments, such as positive, negative or neutral expressions. However, human emotions, such as depression, are very critical and sometimes almost impossible to analyze using these approaches. In this work, we deployed Long Short Term Memory (LSTM) Deep Recurrent Network for depression analysis on Bangla social media data. We created a small dataset of Bangla tweets and stratified it. In this paper, we have shown the effects of hyper-parameter tuning and how it can be helpful for depression analysis on a small Bangla social media dataset. The result shows that 5 layered LSTM of size 128 with batch size 25, learning rate 0.0001 over 20 epochs, the depression detection accuracy is high for stratified dataset with repeated sampling. This result will help psychologists and other researchers to detect depression of individuals from their social activities in virtual world and help them to take necessary measures to prevent undesirable doings resulted from depression.

Nowadays, micro-blogging sites like Twitter, Facebook, YouTube, etc., have become much popular for social interactions. People are expressing their depression over social media, which can be analyzed to identify causes behind their depression. Most of the researches on emotion and depression analysis are based on questionnaires and academic interviews in non-Bengali languages, especially English. These traditional methods are not always suitable for detecting human depression. In this paper, we introduced Gated Recurrent Neural Network based depression analysis approach on Bangla social media data. We collected Bangla data from Twitter, Facebook and other sources. We selected four hyper-parameters, namely, number of Gated Recurrent Unit (GRU) layers, layer size, batch size and number of epochs, and presented step by step tuning for these Hyper-parameters. The results show the effects of these tuning steps and how the steps can be beneficial in configuring GRU models for gaining high accuracy on a significantly smaller data set. This will help psychologists and concerned authorities of society detect depression among Bangla speaking social media users. It will also help researchers to implement Natural Language Processing tasks with Deep Learning methods.

Recently many research works have been conducted and published regarding fractional order differential equations. There are several approaches available for numerical approximations of the solution of fractional order diffusion equations. Spectral collocation method based on Lagrange’s basis polynomials to approximate numerical solutions of one-dimensional (1D) space fractional diffusion equations are introduced in this research paper. The proposed form of approximate solution satisfies non-zero Dirichlet’s boundary conditions on both boundaries. Collocation scheme produce a system of first order Ordinary Differential Equations (ODE) from the fractional diffusion equation. We applied this method with four different sets of collocation points to compare their performance.

Multi-objective simplex algorithm computes all efficient extreme points of a multi-objective linear programming (MOLP) problem. For MOLP problem with large number of efficient extreme points, a filtering technique is necessary to present a small representative part of efficient extreme points. For this, we propose a filtering technique among efficient extreme points. In modern agriculture, we have to consider series of conflicting optimization criteria. For this reason, optimal cropping pattern, that is, allocation of land to various crops by making use of limited resources has become major challenges to get optimum return. To overcome this challenge, we illustrate the application of MOLP through a real life oriented problem of agricultural sector of Bangladesh. We also solve this problem by a self developed computer software which is based on multi-objective simplex algorithm and our proposed filtering technique. Dhaka Univ. J. Sci. 64(2): 147-152, 2016 (July)

This paper presents results of three-dimensional hydrodynamics simulations of the flow inside a model inertial fusion energy (IFE) fusion chamber. Turbulence modeling employing the large-eddy simulation approach is used to estimate the gas dynamics, state, and mixing after a sufficiently large number of target ignitions. The rich radiation-flow physics that takes place immediately after the lasers hit the hohlraum is modeled separately using a high-fidelity one-dimensional model, which provides reference conditions for the complex geometry three-dimensional turbulence simulations. The IFE geometry includes optical ports and recirculation openings as well as a duct to evacuate the debris produced after each energy deposition (as a model of a laser shot). Furthermore, a selected set of sensitivity studies are conducted to estimate the effect of uncertainty in radiative properties of the Xenon gas at the prevalent conditions in the chamber. The results provide guidance regarding the turbulence conditions in the chamber, which seem to have entered a decay state immediately before a new shot takes place. Computational estimates of the density variability within the chamber as well as pressure history at the approximate location of the laser optical ports is presented among other turbulence statistics.

An immersed boundary methodology to solve the compressible Navier-Stokes equations around complex geometries in Cartesian fluid dynamics solvers is described. The objective of the new approach is to enable smooth reconstruction of pressure and viscous stresses around the embedded objects without spurious numerical artifacts. A standard level set represents the boundary of the object and defines a fictitious domain into which the flow fields are smoothly extended. Boundary conditions on the surface are enforced by an approach inspired by analytic continuation. Each fluid field is extended independently, constrained only by the boundary condition associated with that field. Unlike most existing methods, no jump conditions or explicit derivation of them from the boundary conditions are required in this approach. Numerical stiffness that arises when the fluid-solid interface is close to grid points of the mesh is addressed by preconditioning. In addition, the embedded geometry technique is coupled with a stable high-order adaptive discretization that is enabled around the object boundary to enhance resolution. The stencils used to transition the order of accuracy of the discretization are derived using the summation-by-parts technique that ensures stability. Applications to shock reflections, shock-ramp interactions, and supersonic and low-Mach number flows over two- and three-dimensional geometries are presented.

Formulation of LPs and IPs is a technique to convert real life decision problems into a mathematical model. This model consists of a linear objective function and a set of linear constraints expressed in the form of a system of equations or inequalities. In this paper, we present formulation from real life problem as an art. We discuss formulation through real life example and solve them using computer techniques AMPL and LINDO. DOI: http://dx.doi.org/10.3329/dujs.v61i2.17068 Dhaka Univ. J. Sci. 61(2): 185-191, 2013 (July)

The primary cause of packoff in wells is improper hole cleaning resulting from aggregated cuttings and particles from the bottom hole. Drill-bit geometry and rotation plays an important role in dictating returning flow paths of drilling mud and thus transportation of debris particles from the bottom hole. This paper presents an early attempt to investigate the integrated effect of mud flow through drill bit and stabilizer slots on well packoff using computational fluid dynamic (CFD) -based particle simulations. A commercially available three-dimensional (3D) flow solver was used to simulate turbulent flow for both stationary and rotating drill bits and two competing stabilizer designs with different flow by area. The motion of debris was tracked using a discrete phase model in a Lagrangian framework. A set of parametric studies was performed to investigate the effect of stabilizer location, particle size, rate of penetration (ROP), etc. on the hydraulic performance of the integrated assembly in terms of cuttings-transport ratio [the ratio of average particle velocity to average annulus velocity (Ct)], particle concentration, velocity profiles, etc. for a range of flow rates. Results from the combined drill-bit stabilizer simulations were reported and design bottlenecks were identified for a legacy stabilizer model. New insights were gained for design modifications to address packoff. Simulation results show that the particles follow the flow path created by the slanted blades of the drill bit initially and then begin a swirling motion before passing through the stabilizer slots. The new design has a larger flow by area, and it straightens the rotating flow and hence offers faster evacuation of cuttings. The flow profile becomes more uniformly distributed and prevents any clustering of particles, thereby significantly reducing non-productive time (NPT) and improving tripability of the bottomhole assembly (BHA) in oil fields. This study provides insight into stabilizer design concepts and the stabilizer’s location as it is combined with a drill bit. The integrated simulation approach adopted here enables operators to predict the motion of formation cuttings using a simpler particle model in a larger system and helps prevent drillstring packoff.