Department of Computer Science
Permanent URI for this communityhttps://repository.cuilahore.edu.pk/handle/123456789/16
Browse
125 results
Search Results
Item Balancing Fertilizer Usage with Real-Time NPK Data: A Precision Agriculture Model for Sustainable Agriculture in Pakistan(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2025) Jawad Hassan; CIIT/SP24-RCS-005/LHR; Dr. Tariq Umer; LHR TO 10036Pakistan is facing a lot of problems in the agriculture sector as the use of imbalanced fertilizers and inefficiency of nutrients management is going to cause the very low yield of the agricultural fields. It is a matter of fact that traditional agricultural methods often lead to the inappropriate distribution of fertilizers that affects the crop yields, hikes soil infertility, and devastates the environment even more. Either more or less fertilizers usage disrupts the nutrient balance in the soil thereby creating longer agricultural inefficiencies. Utilizing real-time soil and climate monitoring and analysis can provide the technological power to deal with these challenges as AI-driven precision agriculture gains popularity. This research is dedicated to the development of an intelligent fertilizer recommendation system by using sensor-based NPK analysis that is combined with Al decision-making. This system will use prompt and accurate data from soil nutrient sensors, weather conditions, and crop requirements to provide recommendations of a proper fertilization dosage to specific fields. The novel approach seeks to minimize fertilizer wastage, to lower the environmental impact, and to maximize crop yields. The new method aims to reduce the waste of fertilizer, to keep the environmental impact at a low level, and to make the harvest bigger. One of the approaches involves the use of IoT-based soil monitoring systems to gather real-time data, training of machine learning models on historical soil and crop data for nutrient requirements prediction, and the development of a user-friendly decision-support system for farmers. This will be achieved through the provision of actional insights for the farmers who will later take the necessary course of action in their fertilizer applications. Optimizing nutrient management through the use of the new system is the way to go to stop the degradation Environmental issues like soil acidification andItem ULIE: Underwater Low-Light Image Enhancement Using Deep Learning(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2025) Muhammad Umer; CIIT/SP24-RCS-018/LHR; Dr. Muhammad Aksam Iftikhar Associate Professor, 55P; LHR TP 10040Among the Underwater images captured in low-light conditions are usually suffer from severe quality degradation related to insufficient illumination, wavelength-dependent colour attenuation, scattering and noise amplification, while strongly affecting the applicability in underwater vision applications. Correcting these problems while preserving computational efficiency is critical in real-time systems and systems with limited resources underwater. This thesis introduces an efficient deep learning based approach for underwater low-light image enhancement to focus on image improvement in terms of illuminating effect, structural details preservation and color fidelity restoration. This thesis presents light-weight enhancement approach to ensure low computational complexity and a composite loss function is designed to help guide the reconstruction in terms of pixel-level reconstruction, structural similarity, perceptual consistency and color-correction. The proposed method is end-to-end trained and tested on the EUVP Dark dataset by using quantitative and qualitative evaluations. Experimental results show that the proposed approach is able to get significant improvement in illumination recovery and reconstruction accuracy, and the corresponding PSNR value of 28.52, SSIM value of 0.8433 and UIQM value of 2.79. Qualitative analysis also shows increased visibility, balanced color restoration and increased detail clarity in severely degraded underwater images. These results indicate that combining an efficient network design and a task-specific formulation of loss is a good solution for underwater low-light image enhancement, especially for real-time deployment scenarios.Item AI for Media Fairness: Detecting Bias in English News(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2025) Fozia Mujahid; CIIT/SP24-RCS-002/LHR; Dr. Abid Sohail Bhutta; LHR TP 10035The digital journalism growth has enabled many more people to have access to a larger amount of information and news; however, this increased access to the internet and news has also brought attention to an increasing number of additional concerns regarding the influence of bias on how news media report, produce and share their content [1]. In this thesis, we explore and research multiple forms of bias found within English language based media and offer an AI-powered solution to identifying and classifying bias within the English language media [2]. The focus of our research is the major forms of media bias – i.e., (i) language bias; (ii) selection bias; (iii) framing bias; (iv) sentiment bias; and (v) ideological bias. The main data source used in this research is a publically available data corpus known as the Media Bias Identification Corpus (MBIC) which consists of labeled media articles written in the English language. We then utilize Natural Language Processing (NLP) techniques to preprocess the text and to obtain features, and subsequently apply several different types of machine learning and deep learning models to classify media biased content [3]. We thoroughly evaluate the performance of all of the trained models using standard performance metrics in order to compare their classification performance, as well as to apply topic modeling to evaluate how media topics are associated with particular forms of bias. Our evaluation of the experimental results found that framing bias and sentiment-based forms of bias were the two most prevalent forms of bias found within media articles written in English. Finally, the machine learning, and more specifically the deep learning and transformer based, models outperformed models for the purpose of detecting media bias in English written articles. The results of our research will aid in developing a fair and transparent approach to digital journalism for media organizations, by providing insight into how the tyItem Fusion of Infrared and 3D Heatmap Volume Representation of Skeleton Data for Human Activity Recognition(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2025) Muhammad Taimoor Akmal; CIIT/FA22-RCS-025/LHR; Dr. M. Aksam Iftikhar; LHR TP 10031Human Activity Recognition (HAR) is a fundamental component of intelligent monitoring systems in healthcare, rehabilitation, and assisted living, where accurate recognition of fine-grained medical actions is often challenged by subtle motion variations, illumination changes, and background clutter. This research investigates a multimodal HAR framework that integrates infrared (IR) video cues with skeleton- based motion representations to improve recognition reliability for medical-condition activities. A curated subset of the NTU RGB+D dataset is constructed using nine medical action classes (A41–A49: sneeze/cough, staggering, falling down, headache, chest pain, back pain, neck pain, nausea/vomiting, and fan self), comprising 1134 paired IR–skeleton samples with a defined training and validation protocol. The skeleton stream adopts PoseConv3D-style 3D heatmap-volume representations and is trained under two encodings (joint-based and limb-based). The IR stream is trained using an R(2+1)D-18 spatio-temporal backbone, and an ablation study demonstrates that full- frame IR training exhibits severe overfitting, whereas skeleton-guided subject-centric cropping improves generalization and reduces computational redundancy. Multimodal integration is performed through score-level fusion on a paired overlap validation subset, yielding improved recognition accuracy compared to single-modality baselines. Experimental results show strong skeleton-only performance and further gains through IR integration, with the best multimodal setting achieving 96.79% Top-1 and 100% Top-5 accuracy on the paired validation subset. The findings confirm that combining robust volumetric pose cues with appropriately processed infrared features improves discrimination among visually and kinematically confusable medical actions, providing an effective direction for multimodal HAR in healthcare-oriented environments.Item Efficient Anomaly Detection in Surveillance Videos Using Pre-Trained CNNs and Self-Supervised Learning(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2028) Rao Abrar Nawaz; CIIT/FA23-RCS-014/LHR; Dr. Muhammad Hasanain Ch; LHR TP 10034The fast growth of surveillance systems in both public and non-public spaces has led to the creation of large amounts of video data, whose manual surveillance becomes inefficient and unreliable. Automated video anomaly detection has thus evolved to be an important topic area in computer vision and that seeks to detect rare, unexpected, and even dangerous events in real time. Nevertheless, most of the current techniques use labelled anomaly data or computationally complex architectures, so they have a limited generalization capability and can only be used in real-world surveillance settings. This thesis presents Efficient-VAD, a self-supervised video anomaly detection model that can reach high accuracy and yet is a real-time model. The suggested system uses the pretrained MobileNetV3-Large network to extract spatial features efficiently and a memory-augmented LSTM autoencoder to classify normal behavioral patterns of space and time using video data that is unlabeled. The distinctiveness of abnormal occurrences by observing patterns that do not conform to the usual representations of normal objects leads to the detection of the abnormal objects by the framework by raising reconstruction errors in cases where abnormal objects are observed to deviate. The comprehensive experiments of the UCF-Crime dataset prove the efficiency of the suggested strategy, with an AUC-ROC of 0.9861 in the severe conditions of class imbalance. Moreover, Efficient-VAD is applicable to real-time and resource-intensive surveillance environments with a low 512 MB memory footprint and a rate of 87.51 frames per second. These findings confirm that the integration of self-supervised learning and memory-enhanced temporal model is effectivItem Formal Modeling and Analysis of Distributed Measurement Systems in Cyber-Physical Systems Using Colored Petri Nets(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2025) Aqsa Nawaz; CIIT/FA23-RCS-002/LHR; Dr. Farooq Ahmad; LHR TP 10033This research addresses the reliability and consistency challenges associated with the Cyber- Physical Systems (CPS) driven Distributed Measurement Systems (DMS), which capture, collect and process DMS data or information from measurement nodes located in various geo- graphical areas. Due to being distributed and asynchronous, these systems suffer from causal inconsistencies and synchronization delays, and unstable scheduling, thereby undermining dependency in fields like power management systems and smart grids. The first stage of this approach aims to obtain a formal specification of the DMS behaviour using Computation Tree Logic (CTL) and Linear Temporal Logic (LTL). These formal languages clearly capture safety and liveness properties like the absence of deadlock, safe power switching, reliable communication, safe and correct data processing, etc. The second stage of the approach is to build a system under study, a hierarchical Coloured Petri Net (CPN) model in CPN Tools1 using well-defined ColSets, functions, transitions, guards, and inter-page communication to simulate measurement power behaviour and control switching with respect to scheduling, measurement flow, and faults. An executable and semantically rich operational model of the system is obtained from this model. The final step validates the value of the CPN model with a detailed state-space analysis. State- space properties, such as boundedness, liveness, fairness, home-state behaviour, and Strongly Connected Component (SCC), are computed and compared to formal specifications. The analysis outcomes demonstrate that the modelled system fulfils all formalisms and performs acceptably from the system operational viewpoint within the analysed state spaceItem Enhancing Automated Question Tagging: Integrating Large Language Models with Traditional Retrieval Methods(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2025) Bilal Hussian; CIIT/SP24-RCS-016/LHR; Dr. Ghulam Rasool; LHR TP 10038When people ask questions on platforms like Stack Overflow or Quora, the system needs to automatically assign relevant tags to help others find and answer those questions quickly. This is called automatic question tagging and is quite essential since it directly influences the efficiency with which the user can receive help. The majority of existing tagging systems just examine the words of the question itself and the tags accessible in the database. They are deprived of expansive real-world information that can guide them in making superior tagging decisions. Large language models such as GPT-4 are very good at context, real-world understanding, and this may be invaluable in this case. The issue is, these models cannot be confident in selecting tags out of a list of predefined tags of a particular site that they may propose tags that do not exist in the system at all. To address this problem, I created EAQTILLM (Enhancing Automated Question Tagging Integrating Large Language Model). My method is in two steps: I will first run a tagging model that will give me a shortlist of possible tags based on the database of the website. I then run the question through a large language model with prompts that are designed to explore the question in more depth and choose the most suitable tags using that shortlist. By doing this we get the advantage of both worlds the ability of the LLM to understand and the restriction of just being able to pick out of valid and existing tags. I applied EAQTILLM (Test) to two real-life datasets, achieving a precision of 0.4952, recall of 0.6667, and accuracy of 0.9842, which demonstrated remarkable performance. The algorithm was always ahead of current state-of-the-art algorithms, demonstrating that the combination of the traditional ones and large language models is a good direction to enhance automatic question tagging in online communitiItem Histone Explorer: AI-Enhanced Identification and Characterization of Histone Modification(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2025) Muhammad Umar Farooq; CIIT/SP24-RCS-017/LHR; Dr. Farooq Ahmad; LHR TP 10039Histone modifications are one of the key epigenetic mechanisms to control the structure of chromatids and gene expression and play a pivotal role in cellular differentiation, development, and pathogenesis. In spite of the fact that Chromatin Immunoprecipitation Sequencing (ChIP-seq) is an experimental methodology that allows identifying histone marks through the use of high-resolution sequencing, it is expensive, labor-intensive and has low scalability, preventing its general use in a wide range of biological studies. To overcome these difficulties, this research suggests to consider Histone Explorer as an AI-enhanced computational system of the correct recognition and description of sites of histone modifications directly on the basis of genomic sequence information. The suggested framework will consist of a bioinformatics preprocessing pipeline that is reproducible, and sophisticated artificial intelligence models. Peaks of histone modification genomic intervals were obtained at the public repositories of epigenomic repositories such as ENCODE and Roadmap Epigenomics and mapped to a reference genome, hg38, to produce a curated FASTA sequence dataset. An extensive feature extraction scheme was used in order to obtain local and global sequence features, such as nucleotide frequency patterns, k-mer representations, positional features, pairwise interaction features, and higher-order statistical moments. It was trained and tested on the following models: CNN, BiLSTM, GRU, and attention- based transformer models among various models trained using machine learning, deep learning, and transformer-based architectures. There were 100,000 positive and negative samples of Histone sequences in the dataset. The performance of the models was strictly tested with independent test sets and using the k-fold cross-validation protocols. The experimental findings show that the transformer-based Histone Explorer model performs better than both classical and deep learning baselines with an accuracy of 91.5, precision of 90.2, recall of 92.8 and F1-score of 91.5 and AUC- ROC of 95.0. Such findings demonstrate that there is high generalization properties and strength over a wide spectrum of genomic regions. All in all, Histone Explorer offers a scalable, precise, and biologically meaningful AI-based framework of histone modifications identification. The proposed system can provide a useful computational resource in epigenetics studies, biomarker discovery, and precision-medicine studies because it lowers reliance on expensive experimental studies and allows analyzing epigenomics on a high-throughput.Item Semantic-Based Summarization Using Ontology and Machine Learning Algorithms(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2025) Tayyaba Amber; CIIT/SP24-RCS-011/LHR; Dr. Adnan Ahmad; LHR TP 10037In natural language processing, the task of text summarization is a challenging and intriguing problem. A summary of a long document enables the reader to easily grasp the key informa- tion without having to read the entire text. The existing methods struggle with generating fluency and coherent summaries, often missing important sentences in a document. This the- sis aims to automate text summarization by using two approaches: extractive and abstractive. In the extractive text summarization, we introduce two layers of sentence scoring. First, we identify different features of the sentences based on a story document and assign weights to them. Second, we readjust the scoring by applying various rules. The top-N sentences are selected as our extractive text summary, where all the sentences have a high score. In an abstractive text summary, we utilize a predefined T5 model, comparing it with Bert2Bert, PEAGAUS, BART, and GPT2 models. The summary of the document is created based on the level of importance of the sentences in the document. The results obtained using the T5 on the feature-extraction with rules refinement scoring dataset are competitive with state-of-the-art. For the FER model, the performance has been improved, among which the most significant improvements are: BERTScoring 0.88, 0.80, 0.83, for the three selected story documents. After these utilization we also create a FER ontology their each sentences is work as instance and their score are work as its data properties. By using this technique we make our model to query-based user intension summary generator. This ontology give the same result as we pass a query that calculate all the feature and give the high score sentences as an output. By using this technique user also get the summary as their need.Item Advancing Pediatric Wrist Fracture Detection with YOLOv12: Exploiting Attention Mechanisms and Transfer Learning(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2025) Maham Nasir; CIIT/SP24-RCS-020/LHR; Dr. Aksam Iftikhar, 69 P; LHR TP 10041Pediatric wrist fractures are among the common musculoskeletal injuries that have been experienced in emergency and radiology departments. X-ray images diagnosis would be crucial to preventing long-term effects of malunion, chronic pain, and functional impairment. Nevertheless, the pediatric X-rays of the wrist are not able to easily identify fractures because of fine fracture fissures, growth plates, and structural fusion, as well as, the inconsistency in image quality. Even though methods based on deep learning have shown good performance in fracture detection, most of the available methods are confined to classification tasks or fail to perform sufficiently sensitive and accurate localization on small childhood fractures. In this thesis, an improved YOLOv12-based attention-enhanced model is proposed to detect and localize pediatric wrist fractures in X-ray images in an automated manner. The suggested strategy takes advantage of the newly introduced YOLOv12 with its Area Attention (A2) modules, Residual Efficient Layer Aggregation Networks (R-ELAN), and transfer learning that enhance the discrimination of features and multi-scale representation. To ensure the quality of X-ray images and make the models more resistant to exposure, contrast, and anatomy changes, a complex medical-specific preprocessing and data-augmentation pipeline is used. The model is trained and assessed with the GRAZPEDWRI-DX publicly accessible dataset that comprises of more than 20,000 pediatric wrist X-ray images with accurate annotations. Massive experimental tests are performed based on various input resolutions and variants of YOLO, and the performance is measured in terms of precision, recall, mean Average Precision (mAP50 and mAP50-95), and the confusion matrix. The proposed model performs much better than state-of-the-art YOLO-based methods with a precision of 0.797 and a recall of 0.565, mAP at 50 of 0.669, and mAP at 50-95 of 0.381