Department of Computer Science
Permanent URI for this communityhttps://repository.cuilahore.edu.pk/handle/123456789/16
Browse
11 results
Search Results
Item Vision Based Rehabilitation and Information System for Patients(Library Information Services, COMSATS University, Lahore Campus, 2019-06-19) Syyam Noor, Muhammad Hasham; FA15-BCS-127 , FA15-BCS-058; Dr. Allah Bux Sargano; LHR TP 5898The system is an action verification system for rehabilitation of patients with limited mobility. In this regard, a Kinect depth sensor is used to capture the image-sequences of patients for identification of different actions in real time during rehabilitation exercisers at home. The system is trained to make decisions whether the person is performing the rehabilitation exercises correctly or not, each exercise contain some passing parameters, and this is to ensure that all the exercises are done in a complete and perfect manner. A desktop based application is designed for user’s interaction. This application contains a list of exercises available, the user will have an option to choose any of them according to the need. Physiotherapists using this system at their clinics will be hugely benefitted with this, it will act as their personal application based assistant for mobility exercises. Especially, the report generation feature will help them in tracking the patient’s progress. The report can be generated at any given time while doing an exercise that will brief the user about the angles and distances between the joints and what is the ideal one. The main ambition behind making this rehabilitation system is to recover pre-injury stages in all phases of physical fitness. The equipment required by the system are economical ones to help those patients who genuinely need this system which everyone can set up at their own places. This would allow patients to perform the rehabilitation exercises in sitting position at their relaxing place while preventing patients from falling down during training. The proposed system is 100% real-time (gives results on run time), effective to vision-based action identification, and demands low-cost system requirement (hardware and software), which is affordable for almost everyone.Item Magnification Independent Approach to Diagnose the Breast Cancer from Histopathological Images(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2022) SHEHROZ TARIQ; SP20-RCS-017; LHR TP 8354; Dr. Allah Bux SarganoBreast cancer is one of the top cancers that cause death globally. Hematoxylin and Eosin-stained images diagnose biopsy tissue, and experts are usually upset with the closing opinion. Computer-assisted diagnosis techniques enable to cut costs and improve efficiency in this practice. Automation of breast cancer multi class classification from microscopic images has a significant impact on computer-assisted breast cancer identification or prediction. The purpose of breast cancer multiclass classification is to classify different subtypes of breast cancer (Papillary, Adenosis, Mucinous, etc.) after identifying the Benign or Malignant Class. Yet, multi-class classification of breast cancer from microscopic imageries aspects dual major encounters: First, the excessive problems in methods of breast cancer multi-class classification compared to binary class classification (benign vs. malignant), and second, the slight variances in several classes because of the high-quality image inconsistency forms, high cohesiveness of malignant cells, and inconsistency of color distribution. As a result, although the automatic multi-class classification of breast cancer from microscopic images has considerable clinical importance, it has never been investigated. As we are dealing with microscopic imaging, magnification plays a vital role while classifying cancer. Existing literature techniques exclusively emphasize magnification-dependent binary or multi class classification and do not continue the effort for magnification-independent breast cancer diagnosis. Using a proposed model inspired by Vision Transformer (ViT) model with multiple variations, this work offers a robust and novel breast cancer multi-class classification approach combined with a binary classification method from a clinical standpoint. The evaluation methods that are utilized in our work are the accuracy score and confusion matrix. The comprehensive experiments are conducted using the publicly accessible benchmark breast cancer dataset named as BreaKHis. As we have trained three models; first for binary classification and the other two for multiclass classification. We achieved 89.4% accuracy for our first model, 74.57%, and x 57.41% accuracies for the remaining two multi-class models. These outcomes are astonishing as related to existing literature. As existing literature technique claims 88.9%, 63.6%, and 52.7% results against these models. The outcomes showed that the proposed methodology surpassed existing methods in breast cancer diagnosis by a significant margin. This is because of the Convolutional layers, we added before passing to the transformer in our proposed model. This convolution allows the features of the breast legion to be further prominent before applying the multi-head attentionItem A Robust Multi-class Skin Cancer Classification using an Attention-based CNN Model(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2022) Sara Younas; FA20-RCS-015; LHR TP 8352; Dr. Allah Bux SarganoAmong lesions having high mortality rate, skin lesion is at top of the list. The patient‟s life could be rescued through early diagnosis. The manual identification period could be prolonged, as different categories of skin cancer have high similarities between their structure, color, and size which could lead to misclassification. An automatic and robust system is crucial for the timely and accurate categorization of skin cancer. In the last few decades, deep learning emerges as revolutionizing field, especially in the area of medical imaging. Most state-of-the-art work adopted transfer learning and ensemble learning based techniques for this problem, in which models used for TL and EL are designed for different types of problems and mostly trained on the huge amount of data which sometimes don‟t perform well for distinctive skin cancer problems due to very challenging datasets. This work proposed novel architecture based on a deep learning technique, which is designed especially for skin cancer classification problem. For learning, this method used dermoscopic images, which are excellent quality images captured through a high-level magnifying device, to obtain clear insights of skin containing cancer cells. The proposed architecture‟s main contribution includes, enhancing performance by embedding inception residual (IR) blocks in which, inception block is warily designed with several parallel layers that are merged together. Moreover, residual connections are established in each block to learn from low and high-level features concomitantly. These connections accelerate the learning process along with coping with the vanishing gradient problem. Deployment of the attention unit in the network boosted performance by suppressing the value of noise-containing features and assigning maximum value to important and relevant features during the learning process. Reducing the number of parameters by keenly optimizing the size and no of filters is another milestone attained through suggested model. The presented architecture achieved promising results by attaining 91.63% accuracy, 91.60% x precision, 91.60% sensitivity, 91.52% specificity, and 91.60% f1-score on the ISIC-19 dataset.Item Vehicle Detection for Traffic Monitoring from Urban Video Surveillance Cameras using Deep Learning(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2022) Ather Rafiq; FA19-RCS-032; LHR TP 8344; Dr. Allah Bux SarganoIn recent years, vehicle detection for traffic monitoring from urban video surveillance cameras has become a hot research topic among researchers because of an increase in anomalous or unusual vehicle activities from video sequences captured from the traffic surveillance cameras. Instead of manually analyzing the video for detection of anomalies, there is a need for an automatic process that would easily be easily applied to a large number of videos, because the number of video surveillance cameras is increasing in the public places causing the increase in automated analysis of traffic by capturing videos. Therefore, automatic video surveillance of traffic is considered one of its main applications. The main purpose of the video-based surveillance system is to analyze patterns and behavior, vehicle tracking, detection of anomalies, and abnormal event prediction. In this research work, a novel framework: Vehicle Detection for Traffic Monitoring from Urban Video Surveillance Camera (VDTMUVSC) using deep neural networks is proposed to get better results as compared to other state-of-the-art methods which are being used for automobile detection. In this method, to reduce the time for training, pre-trained weights are used in terms of transfer learning and some initial layers from the backbone of architecture are frozen. In the second part, the hyper-parameter tuning technique is used to achieve higher accuracy. Further, extensive experiments have been conducted on the benchmark dataset UA-DETRAC which is introduced recently, especially for the purpose of vehicle detection and tracking. The results demonstrated that our proposed architecture outperformed existing techniques with a margin of 3% to 5% in object detection for vehicles, achieving 80.3% mean average precisioItem Person Re-identification for Real-World Scenarios using Deep Learning(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2023) Muhammad Shahzaib Minhas; FA21-RCS-027; LHR TP 8680; Dr. Allah Bux SarganoWith the increasing need for efficient monitoring and tracking systems in surveillance, Person Re-identification (Re-ID) has emerged as a pivotal solution for monitoring public spaces and automatically tracking suspicious individuals. While significant advancements have been made in the person Re-ID domain, the challenge of continuously evolving new data remains a critical concern for practical model deployment. Most existing works rely on offline learning, which poses difficulties in adapting to dynamic environments, accommodating evolving identities, and ensuring efficient resource utilization in real world scenarios. To address these limitations, there is a pressing need for an active learning system capable of being incrementally updated with new data. In this research, titled "Person Re-identification for Real-World Scenarios using Deep Learning," we propose an innovative framework for iterative model training on progressively evolving data. We selected two benchmark datasets, Market-1501 and DukeMTMC-reID, and divided them into manageable chunks. Our approach involved training a deep learning model incrementally on these chunks. Following each training session, we updated a Representative Memory with the most characteristic images of each identity from each chunk. These images, stored in the Representative Memory, were then utilized in subsequent model updates, where they were combined with new data for further training. To evaluate the efficacy of our model, we conducted a cross-domain evaluation after each iterative training phase. The results were promising with the model achieving accuracy rates of 77.2% and 68.9% in terms of Rank-5 metrics on the Market-1501 and DukeMTMC-reID test sets, respectively. This research not only demonstrates the feasibility of incremental learning in person Re-ID but also paves the way for future advancements in applying deep learning techniques to real-world surveillance and security applications.Item A Machine Learning Approach for Mutation Identification in Gene Sequences to Detect Lungs Cancer(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2022) Umer Farooq; SP19-RCS-029; LHR TP 8334; Dr. Allah Bux SarganoCancer is one of the leading causes of death worldwide. The mutation is one of the basic causes of cancer, which is caused by any change in the gene sequences of the human body. The abnormal and uncontrolled growth in the cells of the lungs is a cause of lung cancer. It is the second most common type of cancer worldwide, considered 11.2% of all cancers. According to the cancer statistics of the United States of America (USA), about 236,740 people were diagnosed with lung cancer in the year 2021, and about 130,180 people died from lung cancer in the same year. This study aims to develop a dataset, machine learning, and deep learning-based techniques to address the above-mentioned issues. In this regard, the asia.ensembl.org website is used to extract normal gene sequences from the lung cancer dataset, while the mutated gene sequences for lung cancer are obtained from the intogen.org website. Thus, the dataset of 1588 samples consisting of 2576 mutations is developed by following the standard dataset creation method. Then, different feature extraction methods are applied to extract the discriminative features from the dataset; these features include raw moment, Hahn moment, central moment, position relative incident matrix (PRIM), reverse position relative incident matrix (RPRIM) calculation, feature vector determination, accumulative absolute position incidence vector (AAPIV), reverse accumulative absolute position incidence vector (RAAPIV). These features are then fed to the traditional machine learning (ML) algorithms such as Random Forest (RF), Logistic Regression (LR), Multi-Layer Perceptron (MLP), and an ensemble of these traditional ML algorithms. These extracted features are also fed to deep learning (DL) methods such as LSTM, GRU, Bi-LSTM, and an ensemble of these DL algorithms. An averaging technique is applied to ensemble the classification algorithms. Three validation methods are used for evaluation: self consistency testing, 10-fold cross-validation, and independent set testing. The final classification result of lung cancer is generated as accuracy, specificity, Mathew's Correlation Coefficient (MCC), sensitivity, and Cohen's Kappa. The efficacy of this study can be measured with the accuracy of 94.5% obtained from the ensemble of x machine learning models using 10-FCV and with the accuracy of 97.3% obtained from the ensemble of deep learning models using the independent set test.Item Activity Recognition for Assisted Living based on Multimodal Features using Deep Learning(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2021) Hafsa Yaseen; FA19-RCS-022; LHR TP 7599; Dr. Allah Bux SarganoHuman activity recognition (HAR) is a prominent field in computer vision and signal processing that analyzes the information obtained from numerous sensors, including vision sensors and wearable sensors. The purpose of HAR is to recognize actions from a sequence of observations on the activities of individuals and environment events. It provides a broad variety of applications, including ambient assisted living, robotic technology, intelligent surveillance, human-computer interaction, smart home, transportation, and smart healthcare. As assisted living indicates the technological services that help impaired people and senior citizens to spend independent life. Therefore, HAR, which facilitates proactive gestures and interactions with their surroundings, has become a significant precondition for assisted living applications. For all that, tremendous efforts have been made to reliably capture human action and behavior by manipulating single modality data, but the combined analysis of multimodal data has received less attention. Different modalities usually contain complementary information that must be combined for better learning of action recognition for Ambient Assisted Living. In this research, a novel framework called “Activity Recognition for Assisted Living based on Multimodal Features using Deep learning” is proposed to leverage intra-modality discriminative features as well as inter-modality connection in visual and inertial data using deep neural networks. Two separate unimodal, i.e., visual and inertial models, are proposed to learn action recognition classifiers for these modalities effectively. These models automatically acquire high-quality discriminative action-related images and inertial features. Finally, these heterogeneous models are combined into an end-to-end approach via decision-level fusion. The comprehensive experiments are conducted using the publicly accessible benchmark C-MHAD dataset. The outcomes showed that the proposed methodology surpassed existing methods in action recognition by a significant margin, with an F1-score of 89%Item A Lightweight Indoor Smoke Detector with Benchmark Dataset using Deep Learning(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2021) Raheel Shahzad; FA19-RCS-007; LHR TP 7602; Dr. Allah Bux SarganoA fire disaster is one of the most dangerous events that may occur at any place and time. To avoid such incidents, smoke detection is crucial before getting fire. The smoke sensors are the most widely used devices to detect fires; however, smoke sensors can only detect fires if the fire is large, and smoke reaches the sensor. Therefore, vision based smoke detectors have been proposed using machine learning and deep learning based methods. Most of these methods have been proposed for outdoor smoke detection, while little attention has been paid to indoor smoke detection due to the lack of appropriate datasets for indoor scenarios. This study creates a benchmark dataset for indoor smoke detection by properly following annotation criteria outlined by Inter Annotator Agreement (IAA) and Cohen’s Kappa evaluation metrics. The proposed dataset achieved 0.91 IAA and 0.81 Cohen’s Kappa scores, which confirms the excellent quality of the dataset. In addition to this, an innovative transfer learning-based method has been proposed for indoor smoke detection and evaluated on the proposed dataset. Furthermore, a state-of-the-art smoke detection algorithm has been implemented and evaluated on the proposed dataset for comparative analysis. The results confirm that the proposed method outperforms the state-of-the-art methods.Item Human Activity Recognition Based on Multimodal Sensor Data Fusion using Deep Neural Networks(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2021) Saad Ahmed; FA18-RCS-027; LHR TP 7588; Dr. Allah Bux SarganoIn recent years, Human Activity Recognition (HAR) has been one of the core research areas due to its various applications. It has been attracting growing attention in the computer vision field. In HAR, activities are normally represented using numerous sensor modalities, like vision, inertial, skeleton, audio, etc. However, there are limitations associated with these sensors like Local barriers, image barriers, sensor unreliability, and consumer concerns. Multimodal Human Activity Recognition (MMHAR) solves these problems by using more than one modality sensor to enable the complementary information of different domains in the recognition task. Recently, various Deep learning-based approaches have been proposed for MMHAR and have achieved state-of-the-results. Though great efforts have been made in this area using various modalities, little attention has been paid to the analysis of the dominance and relevance of one modality over another. This research work highlights the importance of multimodal-based sensor fusion using deep neural networks to achieve HAR and highlights which modality has more importance in recognizing activities. This research work proposes a novel deep multimodal fusion network based on two-stream architecture. One model stream uses the Three-dimensional Convolutional Neural Network (3D-CONV) to handle the depth sensor data. At the same time, the second stream of the model uses a Two-dimensional Convolutional Neural Network and Long Short Term Memory (2D-CONVLSTM) for handling the inertial sensor data. Both streams capture features from data generated through depth and inertial sensors. Decision level fusion combines the results generated from both streams to get the final prediction. The proposed model has been evaluated on a publicly available benchmark dataset Berkley MHAD (Multimodal Human Action Detection) and has produced state-of-the-art accuracy of 99.73%, outperforming the previous methods. The depth sensor and inertial sensor data are passed to the proposed model streams separately in a Single modality-based Human Activity Recognition (SHAR) process. It is observed that the depth camera sensor achieves a higher accuracy result of 98.89 % than an inertial sensor accuracy score of 89.34%. x Hence, it is concluded that the depth camera sensor has more importance in the recognition task than inertial sensor datItem Super-Resolution of Medical Images using CNN(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2021) Maha Anis; SP19-RCS-023; LHR TP 7288; Dr. Allah Bux SarganoSuper-resolution is a technique that improves the quality of an image. Super-resolution images contain more detail because they have a high density of pixels. Super-resolution images provide complementary detail for clinical diagnosis, astronomy, and biometry. The medical images should be of good resolution for perceiving fine-grained details in the image so that the patients are diagnosed at their early stages of tumours. If the images have low resolution, doctors cannot diagnose the tumours or segment the malignant regions. For super-resolution images, the cost in terms of machines and the health of patients also matter. If patients are at high risk of cancerous cells, continuous radiation exposure will lead to activation of these cancerous cells. So the goal is to enhance the low-quality scans. This thesis emphasises single image super-resolution (SISR). Previously used techniques for the task of super-resolution were interpolation and reconstruction. However, these techniques have limitations in the loss of information in images like edges and boundaries. Recently, deep learning was introduced and aimed to features extraction from the related images for enhancement automatically. For this purpose, Convolutional Neural Networks (CNNs) are employed due to their excellent results in similar domains. These CNN based approaches achieved encouraging results. Transfer learning have been successfully used for image recognition and human action detection. This research work foremostly in-depth examined the learned representation models incorporating transfer learning and fusion of different models. EDSR and WDSR are the winners of NTIRE challenges for the super resolution of natural images. These models, including SRGAN, are evaluated with and without transfer learning. The proposed concepts unleash the integrated potential of these three models in the medical domain with higher PSNR and SSIM as compared to contemporary models. The transfer learning EDSR, WDSR, and SRGAN achieved 40.38, 39.02, and 38.15 PSNR and 0.964, 0.963, and 0.953 SSIM respectively. The fusion of these three models shows PSNR, and SSIM 37.84 and 0.964 respectively. The results report that proposed approaches outperformed state-of-the-art methods by learning super resolution features on NAMIC datasets. The results highlight the importance of transfer learning