Department of Computer Science
Permanent URI for this communityhttps://repository.cuilahore.edu.pk/handle/123456789/16
Browse
6 results
Search Results
Item Enhancing Brain Tumor Diagnosis through Deep Feature Learning and Spatial Context Awareness in MRI Scans(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2025) Qasim Ali; FA22-RCS-022; Dr. Muhammad Aksam Iftikhar; LHR TP 9488Brain tumor classification is a critical task in medical diagnostics, where accurate and timely detection significantly influences treatment planning and patient outcomes. In this study, a hybrid deep learning framework which joins Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) is developed to classify brain tumors based on the BRATS 2021 dataset using the FLAIR modality. The CNN part learns local patterns (e.g. edges and textures) while the ViT part captures global dependencies over the whole images, which provides insights of the tumor regions in a holistic way. After performing pre-processing on the dataset, which comprised of 53,610 images, the images were partitioned into tumor and non-tumor groups. The training, validation, and test sets were created with 80%, 10%, and 10%, respectively. The listed model performed two independently trained models for 100 epochs at first, then is applied a feature fusion technique which exploits the complementarity of both architectures. The hybrid model was assessed based on conventional metrics like precision, recall, F1-score, and accuracy. The result confirmed that the hybrid method achieved better performance than the individual CNN and ViT models, with improvements in precision and strengthened classification capability. The study emphasizes how hybrid deep learning techniques can improve clinical applications by alleviating challenges related to data variability, noise, and image complexity. The proposed method combining local and global feature extraction methods ensures valuable typing providing a reliable and scalable implementation of automatic brain tumor classification helping improve the accuracy and efficiency of healthcare related problem solutions and medical diagnostics.Item Wildfire Detection using Satellite Imagery(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2025) Faisal Saqib; FA22-RCS-007; Dr. Zeeshan Gillani; LHR TP 9494This thesis presents an innovative approach to wildfire detection using satellite imagery, aimed at enhancing detection accuracy and robustness through the fine-tuning of pre trained state-of-the-art deep learning models. Utilizing a dedicated wildfire satellite imagery dataset, our research focused on optimizing these models to better generalize across diverse environmental conditions encountered in satellite data. Key evaluation metrics employed in this study included Dice Loss, Intersection over Union (IoU), and Accuracy, which facilitated comprehensive performance assessments. Our method involved adjusting existing model architectures to improve their efficacy in segmenting and identifying wildfire-affected areas accurately. The results of our experiments indicate significant improvements in model performance, with increased precision in detecting and delineating wildfire boundaries. These findings not only demonstrate the potential of advanced computational techniques in the field of remote sensing for wildfire detection but also provide valuable insights for future applications in environmental monitoring and disaster management. This work underscores the effectiveness of tailored deep learning solutions in enhancing the capabilities of satellite imagery analysis for critical real-world challenges.Item Self-Attention Based Multi-Model Fusion Mechanism for Diabetic Retinopathy Grading(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2024) Sana Naseem; FA22-RCS-023; Dr. Muhammad Aksam Iftikhar; LHR TP 9332Among the five basic human senses, vision is the most versatile, which depends primarily on the flawless working of human eyes. An eye condition called diabetic retinopathy (DR) is caused by diabetes. Retinal lesions are important signs of DR, and they can be used for early diagnosis. This research will identify and evaluate any pathological alterations brought on by the development of DR. In this research, an attention based deep learning method is proposed to classify Diabetic Retinopathy (DR), wherein CLAHE preprocessing is first used to enhance the image contrast. The publicly accessible dataset APTOS 2019 is utilized in the experiments to evaluate the effectiveness of proposed methodology. However, this dataset has imbalanced class distribution, which was balanced using different data augmentation strategies. For effective feature learning, four pre-trained models namely Inception v3, VGG-19, ResNet-50, and DensNet-121 are modified by adding self- attention blocks following each convolutional base of the pre-trained models. These self-attention blocks enable the model to selectively focus on important parts of DR lesions. Subsequently, the extracted features are fused using late fusion, and a Multi- Layer Perceptron (MLP) classifier is employed to categorize DR stages. The proposed method was validated by performing ablation studies for each pre-trained model. Specifically, classification evaluation metrics including precision, recall, F1- score, and accuracy were computed for each pre-trained model with and without the attention block. The comparison of results clearly showed the effectiveness of adding the attention blocks to each model. Further, the final MLP classifier trained on the fused convolutional features from all pre-trained models was also evaluated for DR grading, which resulted into precision, recall, F1-score, and accuracy of 0.83%, 0.81%, 0.82% and 0.90%, respectively. These results were compared with state-of-the-art methods to show that the proposed method outperforms existing methods significantly. Therefore, the proposed method can be a useful tool for providing a secondary opinion to medical practitioners for DR grading and classification.Item Urdu Cyclic Text Reuse Detection using Sentence Transformers(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2024) Areej Waseem; FA22-RCS-002; Dr. Muhammad Sharjeel; LHR TP 9334Detection of cyclic text reuse is an increasingly growing need in the Urdu language with the greater use of digital technologies. The growth of text reuse, which is the process in which the passage from an original document is copied and then reused with slight changes in wording or rewording without proper credit, has become very easy due to the vast availability of digital text and free online machine translation tools. The latest AI-powered machine translation systems make the process of iterative translation of text across multiple languages and back to the origin very easy, thus generating paraphrases which can be reused without due credit. Detection of such text reuse is important for the whole gamut of applications in the journalism domain, near-duplicate document detection, content reuse in blogs, and plagiarism detection. While lots of research has been done on the detection of text reuses in English, there exist only a few manually created corpora in the case of Urdu text reuse detection, which are real and simulated instances of text reuse. And up to now, no huge corpus or research targets the detection of cyclic text reuse in Urdu. This study makes an effort to fill this gap by developing a huge benchmark corpus through an automated approach based on real cases from journalism. The proposed corpus contains 6,000 manually annotated documents as either partially derived, wholly derived, or non-derived. In the paper, more research is conducted on the effectiveness of different sentence-transformer methods such LaBSE,1in detecting cyclic text reuse. Besides, we have fine-tuned the LaBSE, paraphrase-multilingual-mpnet-base-v2, distiluse-base-multilingual-cased-v2, xlm-r-100langs-bert-base-nli-stsb-mean-tokens, xlm-r-distilroberta-base-paraphrase-v1 model to increase its accuracy with the comparison to the baseline of the N-gram approach. The results demonstrate that when the implemented change is applied to Urdu, the cyclic text reuse detection of the given document using the proposed transformer-based approach with the LaBSE yields favorable results with the best F1 score being 0. 61. The results revived the capacity of the model in dealing with other intricate translation tasks as well as increasing its accuracy in expounding the cyclic text reuse. The contributions from this research are two-fold: first, the construction of suitable and large scale corpus corpora for Urdu cyclic text reuse detection, and, second, the analysis of performance of existing and specifically state-of-art sentence transformer techniques such as LaBSE for improving the performance of the methods for this specific task. The techniques and collection employed through this study will provide immense help to the further improvement of cyclical text reuse in Urdu language and can be applied for other similar languages of the world. This thesis, therefore, does not only provide coverage of the various approaches to detecting cyclic text reuses using the sentence-transformer-based methods but also comes with the creation of a stable set of corpus that will be used to analyze the methods. We believe that the outcomes and the materials provided in this research can be the basis for further investigations and realizations testing for text reuse in various languages.Item A Person Reidentification Framework for Multi- Camera Tracking by Using Tracking-by-Detection Paradigm(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2025) Mian Muhammad Abu Bakar; CIIT/FA22-RCS-008/LHR; Prof. Dr. Zulfiqar Habib; LHR TP 9697The research is based on offline video-based reidentification of individuals by tracking them across multiple non-overlapping cameras, aiming to achieve accurate, efficient, and effective video-based person reidentification by extracting discriminative human attributes and robustly utilizing spatio-temporal complementary features in the tracking process. Person reidentification (Re-ID) aims to identify individuals with the same identities across multiple cameras at particular or different instances of time from images or video sequences. It involves applying deep learning algorithms that help to track people’s positions, motion, and direction. Over the past few years, deep learning- based methods have achieved significant success in this domain. These methods can be categorized as deep metric learning, micro-level feature learning, generative adversarial learning, domain adaptation and transfer learning, sequence feature learning, transformer-based methods, and clustering-based methods. Specifically, sequence feature learning, which focuses on fusion of spatio-temporal information from multiple dimensions, and transfer-learning-based methods, which focus on using pre-trained models trained in one domain and fine-tuning these models by providing samples of the target domain. These methods have achieved better experimental performances compared to the other deep learning-based methods, with significant improvements in accuracy. In the context of person reidentification, persistent challenges such as heavy occlusion and similar appearances, which lead to ID switches, that affect the person reidentification and tracking process, still remain an active area of research. To address these issues, a Single-Camera People Tracking mechanism named Redundancy- Elimination Clustering (REC) is introduced. ID switches mostly occur due to background clutter, heavy occlusion, similar appearances of individuals, or when two or more individuals cross each other in a camera view. These ID switches affect the process of reidentification and tracking. As a person cannot be tracked in different locations at the same time, it becomes clear that the possibility of the occurrence of the ID switches is due to the presence of multiple individuals in the same frame. The frames in which there are multiple individuals are termed here as the Redundant Frames, and we employ Redundancy-Elimination Clustering (REC) to resolve this issue through Hierarchical Clustering. To handle ID switches across multiple cameras, we propose Feature Supervised Clustering (FSC), which utilizes the Hungarian Algorithm and assigns unique global IDs to individuals, enabling robust cross-camera association. To correct the incorrect global IDs assigned to the individuals in the previous Feature Supervised Clustering (FSC) stage, either due to very similar appearances or due to extreme occlusion, we introduce Cross-Camera ID Refinement (CCIR). This method effectively assigns unique global IDs to individuals. In the end, some post-processing, like linear interpolation and handling of edge cases, is also performed. The final proposed method achieved IDF1 (95.67), IDP (96.26), IDR (95.13), Precision (96.26) and Recall (95.86), compared to the base methodology which achieved IDF1 (93.72), IDP (91.8), IDR (95.75), Precision (91.83) and Recall (95.64), which shows the robustness of the proposed method.Item Anomaly Prediction and Analysis Using Deep Learning-Based Techniques(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2025) Muhammad Usman Hanif; CIIT/FA22-RCS-018/LHR; Prof. Dr. Zulfiqar Habib; LHR TP 9699In today's world, where public safety is a growing concern, intelligent surveillance systems play a crucial role in identifying abnormal behavior in public places. To build an effective surveillance system, early prediction of abnormal actions is crucial as it can help avoid potentially dangerous situations before they occur. This thesis focused on the development of a deep learning-based anomaly prediction model that can predict future abnormal actions based on partially observed information of actions from surveillance videos. These videos are often captured under real-world environmental conditions such as poor lighting, occlusions, dynamic backgrounds, visual noise, etc., and only partial information of an action is considered for prediction, which makes the task of forecasting more challenging. Although previous studies have shown progress in anomaly recognition, which focuses on recognizing abnormal action only after it occurs completely, rather than predicting future action. In addition, many existing methods rely on spatial information and the generation of image-based next anomaly frames, usually ignoring the importance of temporal information. This limits its effectiveness in real-world scenarios in which understanding the action sequence and progress of actions over time is crucial. To address these limitations, this research proposed a novel deep learning-based hybrid approach that combines spatial and temporal information. The proposed model uses a pre-trained ResNet50 to extract spatial features and BiLSTM with an attention mechanism to learn temporal dependencies throughout the video. The final classification is performed using a fully connected layer with softmax activation function. Extensive experiments are carried out on the real-world challenging UCF-Crime dataset, where the model is evaluated at various observation ratios, such as 20%, 40%, 60%, and 80% of the video, to simulate early prediction scenarios. The results show that the proposed model achieved 81.4% accuracy, significantly outperforming baseline predictive models, even when only a limited portion of the action was observed, confirming their effectiveness in the early prediction of anomalies. In addition, the model achieved 90.3% accuracy when tested on complete video (100% video observation ratio), acting as a conventional anomalous action recognition. Although the anomaly prediction task achieved lower accuracy than anomaly recognition due to limited information available at early stages, the proposed model can make accurate future predictions much earlier. This makes it suitable for time-critical surveillance scenarios, where early predictions of crimes are more valuable. This study contributes to the advancement of proactive surveillance systems, allowing timely responses to anomalous events before they occur