2025

Permanent URI for this collection

https://escholar.umt.edu.pk/handle/123456789/17345

Browse

Now showing 1 - 9 of 9

Anomalous human action recognition and crime event classification using deep learning
(UMT Lahore, 2025) Kainat Aslam
Human behavior can be classified into two categories: criminal and non-criminal. Surveillance cameras record activities in their environment, but doing so by themselves becomes challenging and exausting. By working through these problems, we attempt to create an intelligent surveillance system that can distinguish and identify illegal behaviors and gestures without requiring constant human observation. Over the years, deep learning models have transformed image analysis by utilizing multi-layer neural networks. Depending on their architecture, all of these models facilitate considerable improvements in model performance and accuracy. We put forward a method that integrates the LSTM model with the layers that allowed for the identification of temporal patterns within the frames and captures high-level characteristics for important identification of anomalies on the DenseNet121 model. Using the UCF Crime data set, experiments demonstrated that our method outperformed current state-of-the-art methods (75% AUC) and accuracy of 73%. In order to implement complete testing and thorough model training across various anomalous occurrences, we enhanced the data.
Classification of blood cells and leukemia using transfer learning
(UMT Lahore, 2025) AMMARA ZAFAR
Classification of blood cells and detection of leukemia are critical tasks in medical diagnostics, often requiring expert knowledge and significant time. Traditional methods can be timeconsuming and susceptible to human error, underscoring the want for efficient, automated approaches or a strategy. This research delves into the detailed examination of image classification using deep learning, evaluating multiple CNN architectures to determine the most effective model. The primary objective is to improve classification accuracy while leveraging data augmentation to enhance model generalization. The research addresses key challenges, including overfitting and feature extraction efficiency, to acquire a robust predictive model. The fine-tuned VGG16 model has higher overall performance as compared to VGG19, DenseNet201, InceptionV3, and MobileNetV2, thus, it is the best model for this research. VGG16 model, after augmentation, returned a training accuracy of 95.56%, validation accuracy of 94.04%, and test accuracy of 92.81% as a predictive performance metric, outperforming all the other architectures. All of these metrics- Precision, recall, and F1-score accuracy are achieved as 0.93 and 0.93, 0.93 respectively, which indicates the model is prolifically balancing sensitivity with specificity. Using data augmentation & integration, it made test loss go from 0.1912 to 0.1824, which helps with generalization and stabilization. Specific images, their overall layout, code, and statistics illustrate the effect that data augmentation has on our classification accuracy. According to independent testing, VGG16 outperformed baseline techniques, underlining the importance of architectural selection and strategies for augmentation. For image classification tasks, the proposed VGG16 model performs best, making it a trustworthy and efficient approach.
Culturally tailored serious game design model for smart agriculture
(UMT Lahore, 2025) Sehrish Riaz
Computerized learning environments have gained increasing attention in recent years, particularly with the widespread adoption of mobile devices such as smartphones and tablets. This research presents a culturally tailored serious game design model for smart agriculture, aiming to bridge the gap between educational value and cultural relevance. The proposed model integrates sociocultural theory and gamification principles to enhance learning effectiveness, engagement, and usability within specific regional contexts. The Smart Agriculture Game (SAG), developed based on this model, incorporates cultural elements such as traditional farming practices, local crops, and community customs. The game design process followed systematic stages ranging from low- to high-fidelity prototyping and validation supported by specialized heuristics for evaluating usability in culturally sensitive environments. SAG was evaluated using standard testing procedures, achieving a high usability score of 90% and showing significant improvement in learning outcomes compared to traditional learning groups. By embedding smart agriculture concepts such as IoT, big data, and machine learning within culturally familiar narratives, the game supports knowledge retention and the sustainable adoption of modern farming practices. The design encompasses key dimensions including agri-elements, educational objectives, technology, usability, sustainability, and cultural context. Requirements were identified through stakeholder surveys and literature reviews to guide interface and module development. Overall, this study contributes a novel model for culturally adaptive educational game design, aligned with the goals of precision agriculture and rural technology adoption.
Deep & machine learning approch for alzheimer’s disease identification
(UMT Lahore, 2025) IRAM SHEHZADI
Alzheimer’s disease (AD) is a chronic, multifaceted brain disease that belongs to the group of dementia worldwide. At the mundane level, a molecular analysis of the pathogenesis of Alzheimer’s and especially at the level of proteins is indispensable for the early diagnosis and, subsequently, for the discovery of treatment s. In this research, we are interested in distinguishing between the Alzheimer’s linked protein sequences and the protein sequences linked to the other forms of dementia. We importantly operate from the GOLD standard dataset, that is a commonly used benchmark set that has a total of 754 protein sequences. This dataset is split into two categories: 304 proteins related to Alzheimer’s disease and 450 protein sequences not related to Alzheimer’s but relate to other forms of dementia. Thus, by sorting out these proteins the study wishes to understand the proteins that are part of Alzheimer’s disease process from proteins that are involved in other neurodegenerative diseases. To realize this classification, the following machine learning and deep learning methods were used: Our strategies involve using some of the most advanced modern learning algorithms, such as CNNs and RNNs, which are powerful in the analysis of numerous biological patterns. We also compared them with other machine learning methods of a more classical nature, such as Random Forest and Support Vector Machines (SVMs) that have been used for classification tasks in bioinformatics. The data was divided into training and a test dataset to train the models and evaluate how well they do under different circumstances. The training dataset was compiled with a view to estimating the parameters of the models while the one used for testing was used for measuring how well the conceived models were able to correctly classify new protein sequences
Identification of fake medical images using deep learning
(UMT Lahore, 2025) OSAMA TARIQ
Technology is advancing so fast that it is harder to capture fake images, and the emergence of generative models has allowed the generation of realistic synthetic images, popularly referred to as deepfakes. Although this progress has opened innovative possibilities, its abuse in critical fields such as healthcare poses great dangers. In this study, the task is to identify deepfake medical images, namely with malignant skin cancer instances. A special dataset was created, which contained not only real dermoscopic images but also synthetic ones, created with Deep Convolutional GAN (DCGAN), Wasserstein GAN (WGAN), and Stable Diffusion. These false pictures look almost similar to the actual malignancies and thus classifying them becomes a problem. A number of deep learning models were tested in equal conditions; VGG16, MobileNetV2, EfficientNetV2B2, and Xception. Nevertheless, the suggested DenseNet121 model, with a modified classification head and trained through transfer learning, demonstrated the highest result on all assessment measures. The images were divided into four classes, 1,200 training, 400 validation and 400 testing images. The models were evaluated on basis of accuracy, precision, recall, F1-score and a confusion matrix. The using of DenseNet121 with a training accuracy of 0.99 in the context of four classes (real and synthetic images using GAN models) showed high generalization and robustness and had the precision of 0.93, recall of 0.92, and F1-score of 0.92. The main limitation of the study is limited computational resources, so it was hard to generate synthetic images, which restricted the volume and variety of the dataset The work is relevant in the area of deepfake detection in medicine and demonstrates the significance of resorting to heavy CNN architectures such as DenseNet121 when constructing safe and trustworthy AI-powered diagnostic pipelines.
Personalization and gamification based first aid training
(UMT Lahore, 2025) Fatima Farooq
The growth in interest has shown great promise in game-based learning, with particular potential found for first aid training. This study illustrates the potential of gamification to address several identified limitations of traditional methods of first aid training. The present study aimed at the design, implementation, and evaluation of a gamified first aid training system using the First Aid Serious Game Design Model (FASGDM) to enhance user motivation, knowledge retention, and practical application of first aid techniques. The iterative design process was integrated with feedback from experts in the refinement of low and high-fidelity prototypes to ensure usability, accuracy, and contextual relevance with an IRR coefficient of 90% which confirmed the effectiveness and reliability of the gamified system. A quasi-experimental study conducted with 40 students, divided into a control group (n=20) who received traditional lessons and an experimental group (n=20) trained using the gamified prototype. Pre-test results showed equal performance in both groups, while post-test results revealed significant improvement in the experimental group's performance, which was confirmed through statistical analysis t=−23.77, df=24, p<0.05. The gamified approach resulted in higher engagement, improved knowledge retention, and greater confidence in applying first aid skills in real-life scenarios.
Prediction of dipeptidyl peptidase-iv inhibitory peptides using deep learning
(UMT Lahore, 2025) AHMED BILAL
Dipeptidyl Peptidase IV (DPP-IV), a critical enzyme in glucose metabolism regulation, has emerged as a promising therapeutic target for Type 2 diabetes mellitus (T2DM). Inhibition of DPP-IV enhances incretin hormone activity, thereby improving insulin secretion. While conventional methods for identifying DPP-IV-inhibiting peptides are time-consuming and costly, computational approaches leveraging machine learning (ML) and deep learning (DL) offer a transformative alternative for accurate prediction of inhibitory peptides. This study employed peptide sequences filtered using the "Amyloid" keyword from UniProt to construct a robust dataset. A comprehensive computational framework was developed, integrating traditional ML algorithms—Support Vector Machines (SVM), Random Forests (RF), and k-Nearest Neighbors (k-NN)—with advanced DL architectures, including Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks, to classify DPP-IV inhibitors and non-inhibitors. Model stability was validated via k-fold cross-validation, while performance was assessed using accuracy, precision, recall, F1-score, and AUC-ROC metrics. DL models, particularly LSTM networks, achieved superior classification accuracy (96%), outperforming traditional ML methods. Visualization techniques such as t-distributed Stochastic Neighbor Embedding (t-SNE) and Uniform Manifold Approximation and Projection (UMAP) confirmed distinct clustering of inhibitory peptides in feature space. Ensemble learning further enhanced classification reliability through generalized predictions across diverse datasets. The proposed framework demonstrates significant methodological advantages over conventional peptide discovery approaches, including automated feature extraction, high precision, and substantial reductions in experimental screening time and cost. Its scalability positions it as a versatile computational tool for identifying therapeutic peptides across multiple targets. This study underscores the efficacy of DL in deciphering critical peptide features for inhibition and highlights the value of multi-network integration and independent validation for reproducible outcomes. By enabling rapid discovery of novel T2DM therapeutics, this scalable system bridges computational and experimental drug discovery. Future efforts will focus on incorporating additional biological data and transfer learning to refine predictive accuracy and broaden applicability.
Qd2crowdnet
(UMT Lahore, 2025) UZAIR IFTIKHAR
Crowd counting and density estimation have gained significant attention in recent years due to their critical role in applications such as public safety monitoring, intelligent transportation systems, urban planning, and event management. Despite considerable progress made through methods based on convolutional neural networks (CNN), accurately estimating crowd densities remains a challenging task, especially in the presence of occlusions, complex backgrounds, perspective distortions, and varying crowd scales. While many existing models achieve high accuracy on benchmark datasets, they often struggle in complex, high-density scenarios due to limitations in architectural design. Moreover, some of these models are computationally heavy, reducing their suitability for real-time or edge deployments. To address these challenges, this thesis proposes QD2CrowdNet (Quality-Enhanced Depthwise-Dilated Crowd Network), an efficient and accurate deep learning architecture for crowd density estimation. The model leverages depthwise separable convolutions for efficiency, a multiscale dilated backend for context-aware feature extraction, and a refinement module for generating high-fidelity density maps. Extensive experiments were conducted on multiple benchmark datasets, including NWPU-Crowd, ShanghaiTech Part A and B, UCF-QNRF, and JHU-CROWD++. Results demonstrate that QD2CrowdNet consistently outperforms several state-ofthe-art methods in terms of mean absolute error (MAE) and mean squared error (MSE), while maintaining a lightweight structure suitable for real-time applications. QD2CrowdNet achieves the following MAE/MSE scores: 72.68/305.03 on NWPU-Crowd, 54.64/88.1 and 6.2/9.7 on ShanghaiTech Part A and B respectively, 76.0/120.13 on UCF-QNRF, and 52.9/210.32 on JHU-CROWD++. The findings of this study highlight the potential of efficient crowd counting architectures for realworld deployment in complex and dynamic environments.
Skin disease identification using transformer and cnn-based pre-trained models
(UMT Lahore, 2025) SYED WAJID ALI
Skin diseases affect millions of people worldwide and pose a significant diagnostic challenge due to their visual similarity and diversity. This study proposes a deep learning-based approach to classifying 23 types of skin diseases by developing and training models that utilize dermoscopic image data, to enhance diagnostic accuracy. Many deep learning models were tested under the same conditions. VGG16, MobileNetV2, EfficientNet-B0, and ViT Transformer. All of the models, EfficientNetB0 has the highest total accuracy, reaching 0.9377. and for others are MobileNetV2(0.92), DenseNet121 (0.664), ResNet50 (0.255), VGG16 (0.7944) and ViT (0.61). Nevertheless, in the proposed EfficientNet-B0 model, we are using a pre-created dataset by Dermnet from Kaggle, which contains 23 classes of skin diseases. The proposed EfficientNetB0, with an accuracy of 93.77%, precision of 96%, and a recall of 93% had an F1-score of 94%, surpassing all other CNN-based and Transformer-based models of architectures that were tested. Preprocessing procedures such as class normalization, data expansion, and imbalance have significantly improved the output and generalization of the model. Experiments on different architectures and adjustments of hyperparameters proved essential to optimizing results. Evaluation using a confusion matrix and ROC curves confirmed the ability of the model to effectively distinguish visually similar skin diseases. This test shows that improves classification performance, supports dermatologists in clinical decision-making, reduces diagnostic subjectivity, and ultimately provides scalable and reliable equipment to contribute to more accurate and timely treatment of skin diseases.

Browse

Browsing 2025 by Title

Results Per Page

Sort Options