Department of Computer Science
Permanent URI for this communityhttps://repository.cuilahore.edu.pk/handle/123456789/16
Browse
11 results
Search Results
Item 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 Human Action Recognition in Low-Resolution Videos Based on Spatio-Temporal Features(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2023) Samia Akram; FA21-RCS-015; LHR TP 8678; Dr. Usama Ijaz BajwaIn recent years, the field of computer vision and pattern recognition has devoted a lot of research focus to the study of human action. Most of the current action recognition research focuses on high-quality videos with clearly apparent actions. Most actions are of limited quality and take place at a distance, making it difficult to identify them. Therefore, the issue of low video quality is still under-researched and difficult to solve in practical implementation. The goal of this study is to establish a deep learning based framework for recognizing human actions in low-quality video by utilizing spatial-temporal features. This research focuses on the use of annotated dataset TinyVIRAT-v2 of action recognition videos recorded in low quality in order to examine the applicability of deep learning architecture for action recognition from videos of low perceptual quality. First, in the proposed framework datasets will be pre-processed and spatial-temporal features will be retrieved using a CNN-based feature extractor. The extracted features will then be fed into a 3D custom model for classification, which is based on resnet50 as backbone network and c3d with a SoftMax layer for multiclass prediction, allowing for its implementation as a real-world practical application. Upon classifying, every testing video is allocated a predicted category, and the cumulative outcomes are evaluated using performance measure F1-Score which is 0.68.Item Multiclass Weeds and Crops Classification and Segmentation using Deep Learning(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2023) Seemab Ayub; FA21-RCS-009; LHR TP 8675; Dr. Usama Ijaz BajwaThe world's population has increased exponentially with the increase in food demand, so it's essential to focus on factors that affect the quality and quantity of food. Some basic natural factors are climate, soil, pests, and weeds. Weeds are unwanted plants in yields that compete with the crops and consume a lot of nutrients that affect the quality and growth of crops. The separation of these toxic weeds is a challenging task because of their color and quantity. Pakistan is an agricultural country, and almost a big part of its economy depends on agriculture. The growth of this department does not match the population growth; the reason behind this is the presence of weeds. Automated technological interventions represent the optimal approach for weed removal. However, notable challenges persist within the technological landscape, including the absence of comprehensive datasets and the need for well-automated systems specifically tailored for crop weed classification and segmentation. Focusing on these issues, most of the researchers are working on different techniques, but the inadequate size of the dataset is still a problem for researchers, so this study performed to generate an efficient large dataset by merging different small datasets on the base to pick only those classes of weed and crops that belong to Pakistan; also an effective system using UNet variations (UNet, Attention UNet, and Attention Residual UNet(ARUNet)) and ResNet where ARUNet has performed very well for segmentation process and especially the Attention mechanism boosted the segmentation process, on the other hand ResNet has performed best for classification process on our dataset (24- classes and 120000 images). The accuracy of 99% achieved for classification process and 95% for segmentation.Item A Robust Attention-based 3D CNN Model for Fire and Smoke Detection from Videos(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2023) Rimsha Shoukat; FA21-RCS-014; LHR TP 8677; Dr. Usama Ijaz BajwaThe unforeseen occurrence of a fire eruption has significant and adverse impacts on all aspects of human existence, including human life, land, industry, forests, and animal populations. The emergence of smoke before the apparent signs of fire makes smoke detection very advantageous in terms of promptly identifying fires. This early detection capability holds significant potential for preserving numerous lives and mitigating the occurrence of catastrophic events and associated losses. The problem of detecting smoke and fire using video-based methods presents significant challenges due to the considerable variability in their color, brightness, and shape. Most prior researchers conducted individual studies focused on the categorization of fire or smoke. Furthermore, video classification studies focused on a 2D approach, which was inadequate due to its inability to capture temporal information, so it was not able to learn the difference between frames. The convolutional neural network (CNN) has demonstrated exceptional performance in several domains, hence establishing itself as the leading approach. This study presents a robust classifier based on a three-dimensional convolutional neural network (3D-CNN) specifically developed to classify events into three distinct classes: fire, smoke, and neutral. The classifier is meant to effectively process video data. 3D models have the capability to acquire spatial and temporal characteristics from video sequences, rendering them appropriate for undertaking such tasks. The proposed 3D-CNN model had a notable test accuracy of 90% and showcased outstanding performance with an accuracy of 96% when subject to cross data [1] that have seven smoke videos. To enhance its capabilities, the model integrates the CBAM (Convolutional Block Attention Module) attention mechanism. The utilization of this approach allows the model to concentrate on significant regions of interest (ROI) throughout the entirety of the frame, hence enhancing its precision in categorizing complex attributes inside video frames. By integrating the attention mechanism, the model attained a remarkable gain of 6% in test accuracy with an accuracy rate of 96% and an F1-score of 98%.Item Multi-Pollutant based Hybrid Framework for Green Smart Cities using Machine Learning Techniques.(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2023) Usama Masood; FA21-RCS-013; LHR TP 8676; Dr. Tariq UmerAchieving a sustainable environment is one of the most emerging issues discussed in the smart cities concept. Due to the rapid population growth in the world, the concentration of greenhouse gases is increasing day by day. A lot of research studies have focused on different techniques and technologies to reduce environmental pollution. To achieve a sustainable environment, it is important to consider the multi-pollutant factors involve in polluting the environment. In MS thesis, we proposed a multi-pollutants-based intelligent hybrid framework using machine learning that considers the multiple sources of pollution in the city environments. The research emphasizes a comprehensive comparisons of machine learning algorithm for predicting the air quality index in smart cities. In this research. The framework considers the concentration of the pollutants in the environment and will make intelligent predictions on their combined effect on the environment as well as the individual groups based on the similar characteristics of gases concerning the guidelines of WHO. The effect of the concentration of multi-pollutants on air quality and water quality will be analyzed. Machine Learning Techniques including Linear Regression, Support Vector Regression, Random Forest Regression, Decision Tree Regression, and Long Short Term Memory (LSTM) will be applied to predict air quality index and for the classification of pollutants. The evaluation measure used in this research is Accuracy, Precision, F1 score and Recall to find the accuracy of these models.Item Design and Development of a Trust-based secure Authentication scheme for Internet of Drones (IoD) Networks(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2023) Fatima Fayyaz; FA21-RCS-004; LHR TP 8674; Dr. Tariq UmerDrone advancement is prevailing in the latest trends in various sectors. The concept and deployment of commercial drones can provide services in different fields of life. They can be very beneficial in terms of serving industrial, agricultural, construction, packet delivery, videography, and health care providing services in managing, securing, and broadcasting operations. Diversification of networks and dynamic conditions due to heterogeneousness in types of drones and their ownership, a trust factor between these drones is a very critical issue in their intercommunication. Attacker can easily capture the data from a un-secured public channel and can misuse it against participants. Internet of Drone network protection is very important and challenging, as it ensures message integrity, message authenticity and secure authorization access. The lack of this trust factor has just left the Internet of Drones (IoD) domain exposed to security and privacy threats. The drone nodes can only establish a connection to other drones if they are considered legitimate or corporate (friendly) nodes, the rest of them will not be able to make any connection as they are non-corporate (adversary) or unauthorized. The proposed trust factor is based on a security model that establishes trust by continuous authentication and monitoring access attempts between nodes located in the corporate cluster for data delivery. Hence, a trust-based network authenticated framework is much needed for overcoming these security issues. For achieving secure data decimation and intercommunication it is highly important to create a trust-based, threat-free, environment for IOD networks and to provide the strength to security framework in vulnerable authentication schemes against different type of attacks i.e.; replay and impersonation attack, this paper analyzes the research gaps and proposes a new framework for authentication schemes. Considering the previous proposed factors and techniques, this research focuses on establishing a new architecture for providing secure authentication among Internet of Drones (IoD) environment. This paper presents the more secure framework by mutual authentication and key exchange protocol in IoD with continuous chained Hash Function. System generates temporary symmetric keys to establish connection between two nodes, hence devices can mutually authenticate and establishes connection in an untraceable manner. For the verification of sender identity, chained-hashed way is presented. Session keys between devices are updated after every transaction to maintain secrecy. The evaluation of cost and effectiveness of the proposed method with previous protocol are computed. The outcomes demonstrated that our techniques are more secure than previous frameworks.Item Multi-label Financial Text Classification: Corpus and Methods(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2023) Laraib Khalid; FA21-RCS-018; LHR TP 8460; Dr. Jawad ShafiMulti-label financial text classification is evolving but widely researched Natural Language Processing (NLP) task of the current decade. In the recent past, several research efforts have been reported for financial text classification for English and other European languages for instance. Moreover, they are performed for single-label classification (i.e., a document/text is labeled with a single label). However, there is a dearth of South Asian languages, particularly Urdu. Which has around 300 million (non)native speakers around the world and whose digital text is increasing day by day. Unfortunately, very less research efforts have been addressed for Urdu NLP, for instance, financial text classification. Therefore, this research work addresses this research gap for the Urdu language by developing standard evaluation resource and applying baseline supervised and deep learning techniques using multi-label (document/text is labeled with multi labels) classifiers on the proposed corpus. Our proposed Urdu Multi-label Financial Text Classification (UML-FTC-23) corpus contains 24,340 documents and contains different genres and text of business domain. The UML-FTC-23 has been manually annotated with 13 labels ensuring detailed and comprehensive classification of the financial documents/text. It is worth noting that each document in our corpus has been assigned a minimum of two labels and a maximum of nine labels to understand the fine-grained analysis of the financial text and to provide the multi-label classification of text. To demonstrate the quality of the proposed corpus, UML-FTC-23 has been evaluated for multi-label financial text classification by using baseline approaches: 1- Bag of Words (BoW) n-gram, and 2- Term Frequency-Inverse Document Frequency (TF-IDF). Additionally, we have used eight different multi-label classifiers on extracted features. The proposed techniques have been evaluated using several multi-label evaluation measures for instance, exact match, hamming loss, and F1. Furthermore, we have used deep learning methods, including CNN, LSTM, GRU, and RNN on UML-FTC-23 corpus. Result demonstrates that the supervised TF-IDF technique with the multi-label binary relevance classifier achieves the best F1 score of 0.89. Moreover, on the best result using deep learning method are with the CNN which produces the 0.81 F1 score. This indicates that the proposed corpus and techniques are of worth and can be used in various research fields. The proposed corpus and resources are freely available for academic research purposes.Item Cross-genre Multi-label Emotion Classification on Mono-lingual and Code-mixed Texts(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2023) Ahmad Mahmood; LHR TP 8472; FA21-RCS-008; Dr. Rao Muhammad Adeel NawabThis research focuses on This research focuses on Emotion Classification (EC) in textual data and explores the problem of Cross-genre Multi-Label Emotion Classification (CGMLEC). EC involves detecting the emotional attitude conveyed by text, while Multi-Label Emotion Classification aims to assign relevant emotion labels that accurately reflect the author's state of mind. In previous studies most of the work has been done using Same-genre Multi-label Emotion Classification [2], [14]–[16]. Whereas there is only one article that solves the problem of CGMLEC, but the language is the same. However, the problem of CGMLEC has not explored using mono-lingual (English tweets) and code-mixed (SMS messages). Secondly, this research study has Developed, Applied, Evaluated, and Compared Classical Machine Learning, Deep Learning and Transfer Learning based methods. Whereas in Transfer Learning methods we have implied 5 distinct sentence transformers (all-distilroberta-v1, all-MiniLM-L12-v2, all-mpnet-base-v2, facebook-dpr-question_encoder-single-nq-base, and LaBSE), Also the Feature Fussion based approach is applied which combines the features of all the transformers and then the Machine Learning models are applied. The proposed transformers-based approach outperforms with an F1 score of 0.3333 using all-distilroberta-v1 transformer. ) in textual data and explores the problem of Cross-genre Multi-Label Emotion Classification (CGMLEC). EC involves detecting the emotional attitude conveyed by text, while Multi-Label Emotion Classification aims to assign relevant emotion labels that accurately reflect the author's state of mind. In previous studies most of the work has been done using Same-genre Multi-label Emotion Classification [2], [14]–[16]. Whereas there is only one article that solves the problem of CGMLEC, but the language is the same. However, the problem of CGMLEC has not explored using mono-lingual (English tweets) and code-mixed (SMS messages). Secondly, this research study has Developed, Applied, Evaluated, and Compared Classical Machine Learning, Deep Learning and Transfer Learning based methods. Whereas in Transfer Learning methods we have implied 5 distinct sentence transformers (all-distilroberta-v1, all-MiniLM-L12-v2, all-mpnet-base-v2, facebook-dpr-question_encoder-single-nq-base, and LaBSE), Also the Feature Fussion based approach is applied which combines the features of all the transformers and then the Machine Learning models are applied. The proposed transformers-based approach outperforms with an F1 score of 0.3333 using all-distilroberta-v1 transformer.Item Classification of MGMT Promoter Methylation Status in Brain Tumor MR Images using Deep Learning(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2023) Muhammad Sohaib Iqbal; FA21-RCS-005; LHR TP 8470; Dr. Usama Ijaz BajwaGlioblastoma is the most dangerous brain tumor type (worst prognosis), having the least survival rate of less than a year. O6-methylguanine-DNA methyltransferase (MGMT) promoter status “methylation” adds a favorable prognosis and is a reliable indication of cancer's chemosensitivity. Therefore, determining the MGMT status of a brain tumor patient is essential for treatment planning. Surgical tissue sample removal is required to conduct genetic testing for tumors. The genetic characterization of the tumors may then take a few weeks to establish after that. Non-invasive diagnosis using MR images plays a significant role in the early detection of the deadly disease by using Magnetic Resonance Imaging (MRI) to detect the tumorous regions using deep learning techniques. The main requirement for detecting MGMT status is obtaining the sub-regions of brain tumor from MRIs. Previous studies have worked on Brain tumor segmentation and MGMT classification separately, few of the studies have worked on segmentation and classification together but have used selective modalities for MGMT Prediction. Since each modality has its own significance and MGMT promoter status lies in the tumorous region, this study proposed a pipeline of segmentation and classification models using stacked multimodalities of MRI scans of the benchmark dataset BraTS2021. The pipeline is divided into two phases. The first phase uses a stack of multiple MRI modalities to segment the brain tumor into sub-region using a 3D Residual U-Net Architecture, and the second phase uses the segmentation model's output (stacked multimodalities tumor voxel) to determine the MGMT promoter status using a 3D ResNet 10 Classifier. With the help of pipeline, classification model received precise information about the tumorous regions which helped in efficient prediction of the MGMT status present in the brain tumor. The proposed pipeline has therapeutic value because it can assist radiologists in diagnosing brain tumors more efficiently and precisely by reducing the subjectivity and variability of human interpretation. The segmentation and classification pipeline for brain tumors x can potentially increase the consistency and objectivity of diagnoses. Furthermore, it can also assist radiologists in predicting the MGMT status without using surgical equipment and can help in treatment decision making. The segmentation phase of the proposed pipeline yielded promising results with average dice scores of 0.81, 0.84, and 0.80 for the tumor core (TC), whole tumor (WT), and enhancing tumor (ET) on validation. The classification phase achieved ROC-AUC score of 0.66 on validation.Item An Automated Diabetic Retinopathy Grading Pipeline based on Deep Learning Approaches(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2023) MUHAMMAD ZAHID AHMAD; FA21-RCS-007; LHR TP 8471; Dr. Muhammad Aksam IftikharEarly detection is essential for efficient disease treatment and prevention since diabetic retinopathy (DR) is one of the main causes of blindness in the globe. This thesis addresses the value of early detection during the DR stages and the part deep learning plays in creating a reliable detection system. The study stresses the value of early DR stage detection to stop disease development and enhance patient outcomes. Early illness detection and classification allow for the implementation of prompt therapies that can halt or delay the disease's course. In addition to increasing the likelihood that visual function will be preserved, this lessens the need for expensive and intrusive therapies, which eases the strain on healthcare systems. To achieve accurate detection of DR, a deep learning approach is employed. Specifically, an ensemble method of two Vision Transformer (VIT) models followed by Support Vector Machine (SVM) is utilized. We have used Kaggle APTOS 2019 dataset. The ensemble model capitalizes on the strengths of multiple VIT models to enhance feature representation and improve classification performance.The experimental results demonstrate the effectiveness of the proposed model, achieving an overall accuracy of 91.14%, precision of 92%, recall of 91%, and an F1 score of 91% on APTOS dataset. These results outperform existing methods and validate the potential of deep learning in DR detection. The developed model showcases the capability of deep learning algorithms to provide reliable and efficient diagnostic support, enabling early intervention and ultimately contributing to the prevention of DR-related visual impairment.