Research Publications
My research areas and interests are vast and diverse, including but not limited to:
Machine Learning, Deep Learning, Computer Vision, Natural Language Processing
For more insights into my work and biography, please visit Google Scholar
Journal Articles:
ECARRNet: AnEfficient LSTM-Based Ensembled Deep Neural Network Architecture for Railway Fault Detection
Accidents due to defective railway lines and derailments are common disasters that are observed frequently in Southeast Asian countries. It is imperative to run proper diagnosis over the detection of such faults to prevent such accidents. However, manual detection of such faults periodically can be both time-consuming and costly. In this paper, we have proposed a Deep Learning (DL)-based algorithm for automatic fault detection in railway tracks, which we termed an Ensembled Convolutional Autoencoder ResNet-based Recurrent Neural Network (ECARRNet). We compared its output with existing DL techniques in the form of several pre-trained DL models to investigate railway tracks and determine whether they are defective or not while considering commonly prevalent faults such as—defects in rails and fasteners. Moreover, we manually collected the images from different railway tracks situated in Bangladesh and made our dataset. After comparing our proposed model with the existing models, we found that our proposed architecture has produced the highest accuracy among all the previously existing state-of-the-art (SOTA) architecture, with an accuracy of 93.28% on the full dataset. Additionally, we split our dataset into two parts having two different types of faults, which are fasteners and rails. We ran the models on those two separate datasets, obtaining accuracies of 98.59% and 92.06% on rail and fastener, respectively. Model explainability techniques like Grad-CAM and LIME were used to validate the result of the models, where our proposed model ECARRNet was seen to correctly classify and detect the regions of faulty railways effectively compared to the previously existing transfer learning models.
Conference Proceedings:
Automated Detection of Malignant Lesions in the Ovary Using Deep Learning Models and XAI
The unrestrained proliferation of cells that are malignant in nature is cancer. In recent times, medical professionals are constantly acquiring enhanced diagnostic and treatment abilities by implementing deep learning models to analyze medical data for better clinical decision, disease diagnosis and drug discovery. A majority of cancers are studied and treated by incorporating these technologies. However, ovarian cancer remains a dilemma as it has inaccurate non-invasive detection procedures and a time consuming, invasive procedure for accurate detection. Thus, in this research, several Convolutional Neural Networks such as LeNet-5, ResNet, VGGNet and GoogLeNet/Inception have been utilized to develop 15 variants and choose a model that accurately detects and identifies ovarian cancer. For effective model training, the dataset OvarianCancer&SubtypesDatasetHistopathology from Mendeley has been used. After constructing a model, we utilized Explainable Artificial Intelligence (XAI) models such as LIME, Integrated Gradients and SHAP to explain the black box outcome of the selected model. For evaluating the performance of the model, Accuracy, Precision, Recall, F1-Score, ROC Curve and AUC have been used. From the evaluation, it was seen that the slightly compact InceptionV3 model with ReLu had the overall best result achieving an average score of 94% across all the performance metrics in the augmented dataset. Lastly for XAI, the three aforementioned XAI have been used for an overall comparative analysis. It is the aim of this research that the contributions of the study will help in achieving a better detection method for ovarian cancer.
Intracranial Brain Hemorrhage Diagnosis and Classification: A Hybrid Approach
Intracranial brain hemorrhage is a very common problem with a high mortality rate and often can be life-threatening if necessary steps cannot be taken on time. Patients with hemorrhagic cases need to undergo a CT scan of the brain and for taking further steps, the scans should be examined immediately. For this purpose, we proposed a CAD system using a hybrid machine-learning approach which will help radiologists to diagnose intracranial hemorrhage in a more robust way. We used VGG16 and VGG19 models for feature extraction and then trained random forest (RF) and multilayer perceptron (MLP) models with these features. For our research, we have collected a CT brain image dataset that contains 2,501 images with five hemorrhage classes: intraventricular, intraparenchymal, subarachnoid, epidural, and subdural. After training our models it resulted in an overall accuracy of 97.24% using the VGG16-MLP model and 97.02% accuracy using the VGG19-MLP model for classifying brain hemorrhage from CT scans images. A comparative result of our best approach vs. the previous best approach (from our reviewed papers) for each hemorrhage class is as follows; epidural: VGG19-MLP (0.97) vs. YOLOv4 (0.98), intraparenchymal: VGG16-MLP (0.95) vs. YOLOv4 (0.95), intraventricular: VGG19-MLP (0.90) vs. DB-RF (0.97), subarachnoid: VGG19-MLP (0.94) vs. DB-RF (0.90), and subdural: VGG16-MLP (1.00) vs. YOLOv4 (0.95).
DeepGrip: Cricket Bowling Delivery Detection with Superior CNN Architectures
Delivery in cricket is the sole action of bowling a cricket ball towards the batsman. The outcome of the ball is immensely pivoted on the grip of the bowler. An instance when whether the ball is going to take a sharp turn or keeps straight through with the arm depends entirely upon the grip. And to the batsmen, the grip of the cricket bowl is one of the biggest enigmas. Without acknowledging the grip of the bowl and having any clue of the behavior of the ball, the mis-hit of a ball is the most likely outcome due to the variety in bowling present in modern-day cricket. The paper proposed a novel strategy to identify the type of delivery from the finger grip of a bowler while the bowler makes a delivery. The main purpose of this research is to utilize the preliminary CNN architecture and the transfer learning models to perfectly classify the grips of bowlers. A new dataset of 5573 images from Real-Time videos in offline mode were prepared for this research, named GRIP DATASET, consisted of grip images of 13 different classes. Hence the preliminary CNN model and the pre-trained transfer learning models - Vgg16, Vgg19, ResNet101, ResNet52, DenseNet, MobileNet, AlexNet, Inception V3, and NasNet were used to train with GRIP DATASET and analyze the outcome of grips. The training and validation accuracies of the models are noteworthy with the maximum validation accuracy of the preliminary model reaching 98.75%. This study is expected to be yet another steppingstone in the use of deep learning for the game of cricket.
Approaching Deep Convolutional Neural Network for Biometric Recognition Based on Fingerprint Database
Fingerprint dataset is one of the most broadly implemented and broadcasted biometrics for the derivation of individual feature identification. Fingerprint dataset performs in multiple approaches, such as applying query by image content techniques, reviewing criminal offenders, surveillance, taking a difficult decision, searching immediately, and anthropological research because of the uniqueness and persistence of the fingerprint dataset. Here in this research signifies an efficient way of identifying two key biological features: blood group and gender distinguish, based on the fingerprint dataset, applying Deep Convolutional Neural Networks (D-CNNs). The proposed model contains a modified approach of D-CNN and is trained and developed on a self-built fingerprint dataset. Thus, the algorithm applied here aims to observe how prominent the model performs for the custom-built dataset. The proposed model of D-CNN approach proved to be an improved technique and reaches an accuracy of around 99.968% based on the fingerprint images by the individuals for the identification of blood group and gender.
Pneumonia Detection by Analyzing Xray Images Using MobileNET, ResNET Architecture and Long Short Term Memory
Pneumonia is a common and deadly disease that has been in existence for quite a while. One of the handy tools to detect its presence in the human body is the use of X-rays. After obtaining the X-ray images it is manually analyzed by professionals to find out whether the patient has pneumonia or not. Analyzing X-ray images by different professionals might breed conclusions that are not exactly similar, this is due to the variation of making judgments differs from people to people. Additionally, different people can make different errors while analyzing X-Ray images as it is a manual process. Two different X-ray images produced for the same patient might produce different results based on the rotation of the patient’s body on each of the X-Ray imaging processes. To make things simpler, for both professionals and non-professionals for analyzing images, we have proposed a method using deep learning techniques in the form of ResNet 101, MobileNet, and LSTM to analyze X-Ray images and detect whether the patient has pneumonia or not. A dataset consisting of 5856 X-ray images has been used and the maximum accuracy noted of our proposed LSTM model is 95.2%.
Smart Security, Food and Environment Monitoring System for Cows using Microcontroller
Health and security for cows have been an alarming concern over the South-Asian countries as many people depend on cows for a living. Cows are used for various purposes. For example: It can be used for milk production, crop farming and meat production. It is evident that the health of cows depend on what they eat and where they eat. Many measures can assess the quality of the food. One such way of providing reliable food is by the measurement of pH and temperature of the food. The environment in which the cows stay has significant effects in the wellbeing of the cows. Foot diseases can occur if the cows stay near its own dung for too long. Sadly, no farm takes any measurements during the provision of food. Moreover, farms neither check the environmental conditions nor do they maintain it regularly. Furthermore, the security of cows inside the farm is very tough to maintain, as constant labor need to keep eyes on the cows. Our system looks after the health of cows by measuring pH and temperature of food. It keeps the environment clean by spraying water on the farm at regular intervals of time and security is kept in check by detecting whether the cows have crossed the gate of the farm. The system does all these processes automatically by using pH, temperature, RFID sensors, Arduino micro-controllers, water pump and a 9V battery to power up the pump.
Thumb Controlled Low-Cost Prosthetic Robotic Arm
Not everyone is gifted with perfect limbs. There are people in this world who lose their precious body parts in accidents or diseases. They are also not up to speed with the staggering development of technology and productivity. To solve this issue, the area of prosthetic limbs has been widely explored since its inception. However, the existing solutions are financially out of the reach for general people of developing countries. In this paper, a low-cost robotic arm has been proposed which is affordable to above or below elbow amputees. Using servo motors, Arduino UNO, end effector etc. to construct the proposed system can enable a person to move, grasp and replace objects that are heavy or require both hands to carry. A prototype of the model has been made to show the efficiency, productivity and cost analysis of the proposed system. The result displays that the system is cost-effective and affordable to the poverty-stricken section of the community.
Smart Health Care Monitoring System By Seven Segment Image Processing Technique
Health care monitoring (HCM) is an innovation to empower observing of patients outside of traditional clinical settings, which may build access to mind and reduction medicinal services conveyance costs. In order to proper recovery, monitoring a patient’s medication and health condition is mandatory. In daily life, it is very common that patients forget about taking their antibiotics on time and for that they have to start their medication course again. Moreover, in the hospital nurses sometimes cannot give the patient, medicine on time and do other checkups because of work load. In terms of emergency patients, we believe a single minute matter. We are proposing a model where patients will be able to set up their medication time. On time they will be alerted to take medicine. Moreover, we are using pulse sensor and body temperature sensor. These sensors will take reading automatically from the patient on time. After that, all these records will be saved in a text file and doctor will be able to check from his chamber anytime he wants through remote wireless connectivity with our system.
Book Chapters:
An Efficient Traffic Sign Recognition Approach Using a Novel Deep Neural Network Selection Architecture
Traffic sign classification is an important aspect of autonomous driving systems. A slight improvement on classification performance can potentially lower the rate of car accidents. In view of this, we propose three different deep convolutional neural networks in a hierarchical pattern, yet not convoluted among themselves for classifying traffic sign. A very popular and reliable traffic sign dataset called GTSRB is used to train our proposed networks. In our work, we present a novel approach to classify images. Furthermore, we modify all three convolutional neural networks over some of the existing neural nets. While modifying the networks, we redesign them based on specific requirements which may also prove handy for other datasets. Along with the new methods, we are able to reduce the computational complexity as well. On top of the new architecture, we achieve a notably higher accuracy in performance of 99.92% surpassing the state-of-the-art performance of 99.81%. In a nutshell, we trained an artificial intelligence (AI) model that learns to chose between two different AI models while classifying an image.
Book Chapter:
Bangla handwritten character recognition with multilayer convolutional neural network, BM Abir, Somania Nur Mahal, Md Saiful Islam, Amitabha Chakrabarty
Handwritten character recognition from a natural image has a large set of difficulties. Bangla handwritten characters are composed of very complex shapes and strokes. Recent development of deep learning approach has strong capabilities to extract high-level feature from a kernel of an image. This paper will demonstrate a novel approach that integrates a multilayer convolutional neural network followed by an inception module and fully connected neural network. The proposed architecture is used to build a system that can recognize Bangla character from different writers with varied handwriting styles. Unlike previous handcrafted feature extraction methods, this CNN-based approach learned more generalized and accurate features from a large-scale training dataset. 1,66,105 training images of Bangla handwritten character of different shapes and strokes have been used to train and evaluate the performance of the model, and thus allows a higher recall rate for the character in an image and outperforms some current methodologies.
Conference Paper:
Detecting sex from handwritten examples, Sourajit Saha, Md Asif Bin Khaled, Md Saiful Islam, Nisha Saha Puja, Mahady Hasan
There are several tasks that human excel at and computers do not and vice-versa. Just until a few years ago computers were as good as a storage for images and videos. However, in the past 6 years with the boon in artificial neural network, labeled data and computation power; machines have started becoming smart at tasks like recognizing images, detecting different objects in images, captioning images, understanding and summarizing videos, detecting semantic actions in videos and so on. Deep learning researchers and practitioners have started demonstrating notable performance of AI(Artificial Intelligence) on many different tasks that pushes the boundaries and as a continuation of that process, we took one specific problem to solve using deep learning that even human can not solve. We have taken Bangla handwritten characters, then trained them applying several deep learning techniques such as Convolutional Neural Network and Recurrent Neural Network to predict the sex of the writer. Consequently, we have got 91.85% accuracy rate and also demonstrated further analysis of the results that we got.
Human identification using wifi signal, Md Nafiul Alam Nipu, Souvik Talukder, Md Saiful Islam, Amitabha Chakrabarty
Prior research has shown that it is possible to identify human by examining the changes on the WiFi spectrum using WiFi signals. Wireless devices fill the air with a spectrum of invisible RF (Radio Frequency) Signals. When human start walking through this spectrum the signal propagates differently for each person as everyone’s gait, body-shape, and walking style is unique and does not matches with another person. In this paper, we propose a system that uses Channel State Information (CSI) System to extract unique features of an individual’s unique walking pattern. While other identification systems (e.g. fingerprint, face recognition) has certain shortcomings (e.g. intrusive, expensive, and inconvenient), the proposed system overcomes these problems. This system can uniquely identify human with an average accuracy of 78% to 97.5% by using Random Forest (RF) and 84% to 95% average accuracy by using Boosted Decision Tree (BDT) algorithm. The authors believe this system can be used in small office or smart home settings.
Low latency image processing of transportation system using parallel processing co-incident multithreading (PPcM), Sanjana Khan Shammi, Samia Sultana, Md Saiful Islam, Amitabha Chakrabarty
Low latency image processing and high FPS (frames per second) is significant for high resolution decision making in many object recognition applications. Reading frames in between processing of a video is too slow and sluggish as the corresponding reading and decoding the frames are done in the main processing thread. Packages, such as imutils provide such off-the-shelf image processing algorithms which apply multi-threading to achieve low latency. However, the algorithms are unable to perform computationally expensive image processing operations. In this paper, we apply a parallel processing technique based on coincident multi-threading to decrease the latency for computationally expensive cases. The technique is evaluated using a prototype of smart car to show that FPS rate is increased and time complexity of algorithms is reduced by an order of n.
Implementation of an optical character reader (ocr) for bengali language, Muhammed Tawfiq Chowdhury, Md Saiful Islam, Baijed Hossain Bipul, Md Khalilur Rhaman
Optical Character Recognition (OCR) is the process of extracting text from an image. The main purpose of an OCR is to make editable documents from existing paper documents or image files. Significant number of algorithms is required to develop an OCR and basically it works in two phases such as character and word detection. In case of a more sophisticated approach, an OCR also works on sentence detection to preserve a document's structure. It has been found that researchers put lots of efforts for developing a Bengali OCR but none of them is completely error free. To take this issue in consideration, the latest 3.03 version of Tesseract OCR engine for Windows operating system is used to develop an OCR for Bengali language. Moreover, 18110 characters and 2617 words are used to make the OCR's library. In this research, ‘Solaimanlipi’ font and 200 input files are used to test the accuracy of OCR. It is found that for clean image files, the accuracy of the software is as high as 97.56%. It is to be noted that accuracy is measured as the percentage of correct characters and words.
Conferences:
Faster Image Compression Technique Based on LZW Algorithm Using GPU Parallel Processing
As web, mobile, desktop and all other applications use image for different purposes, image compression technique has become one of the most important applications in image analysis as well as in computer science. Though image compression is an old concept, it is still a very time-consuming process which has opened a new field of research in image compression. In this paper, we showcase our extensive research works performed on LZW (Lempel-Ziv-Welch) algorithm, a lossless image compression algorithm that yields faster computation time when it is processed with parallel computation, based on the outcome of our research. The improved time efficiency obtained from our experimental result can be perceived in the context of millisecond scale rather than in the macro level of measurement along with keeping the decoded image in lossless format.