Department of Computer Science

Permanent URI for this communityhttps://repository.cuilahore.edu.pk/handle/123456789/16

Browse

Subject: search.filters.subject.Neural Network

Search Results

Now showing 1 - 3 of 3
  • No Thumbnail Available
    Item
    Hashtag Recommendation for Micro Videos Using 3D Convolutional Neural Network
    (Library Information Services, COMSATS University Islamabad, Lahore Campus, 2022) Bilal Ahmed; SP19-RCS-009; LHR TP 8351; Dr. Ashfaq Ahmad
    In recent years, due to common usage of camera equipment like mobile phones and variations of various short videos platform, a lot of videos published each second are either creative or non-creative. Compared to short videos creation, traditional video creation process is very long process like time consuming, producing & casting. Creating shot video is easy where you can use any smart device’s camera, a video is creative if it creates a meaningful interest in your mind after watching else non creative. In this paper we focused on a deep learning algorithm for understanding consistent features and complementary features of micro videos in vine dataset using 3-dimensional convolutional network. The algorithm works on equal-sized frames of video to extract & learn features such as spatial features where we train the model on three different modules of vine dataset d60, d80 & d 100 of vine. We also perform batch normalization on convolutional outputs to avoid overfit & got best results for given vine test data. Through experimental practice we found that 3D CNN performs better than previous methods of understanding video method. In addition to given algorithm we found that how different training dataset affect the feature extraction and affect the results
  • No Thumbnail Available
    Item
    Metropolis Violation Detector Using Deep Neural Network
    (Library Information Services, COMSATS University Islamabad, Lahore Campus, 2021) Iqra Khalid; FA19-RCS-014; LHR TP 7298; Dr. Zeeshan Gillani
    Anomaly discovery on road traffic is a significant assignment because of its extraordinary potential in metropolitan traffic the board and street wellbeing. It's anything but an extremely difficult assignment since the strange occasion happens seldom and shows various practices. In this work, we present a model to recognize abnormality in street traffic by gaining from the vehicle movement designs in two particular yet connected modes, i.e., the static mode and the powerful mode, of the vehicles. The static mode investigation of the vehicles is gained from the foundation demonstrating followed by vehicle identification technique to find the unusual vehicles that keep still out and about. The unique mode investigation of the vehicles is gained from identified and followed vehicle directions to find the strange direction which is atypical from the predominant movement designs. The outcomes from the double mode investigations are finally melded by driven a re identification model to acquire the final abnormality. This study was based on three classes of anomalies (car crash, car stall and lane change). It was divided into a two class problem with 63 anomaly folders and 86 no anomaly folders. Dataset was provided by NVIDIA AI city challenge track 4 containing 100 train and 100 test videos. Firstly videos were annotated according to anomaly event time. The anomaly time was converted into seconds and then specific frames were cropped. Then 30 frames per video were selected to fed into ResNet18 for high feature extraction. Then LSTM architecture and 3D-CNN was trained and results were evaluated on accuracy evaluation measure. Satisfying results was obtained from both architectures i-e; 80 % accuracy of 3D-CNN and 83 % accuracy of ResNet with LSTM. As it is an on-growing field many work can be done in future by using advanced models for the improvements
  • No Thumbnail Available
    Item
    Neural Network-Based Anomaly Detection in Cryptocurrency Transactions
    (Library Information Services, COMSATS University Islamabad, Lahore Campus, 2025) Rashida Noreen; CIIT/SP23-RCS-025/LHR; Dr. Tariq Umer; LHR TP 9705
    Cryptocurrency, especially the rapid growth in adopting bitcoins has led to a significant increase in fraud and illegal transactions, which faces serious challenges for financial integrity and digital trust. Methods of detecting traditional fraud are often reduced to detecting sophisticated patterns within the blockchain network due to decentralized and dynamic nature of transaction flows. This thesis presents a graph-based deep learning approach to detect discrepancy in cryptocurrency transactions, which uses a large, label dataset of bitcoin transactions collected over time. The functioning begins with the collection and prepressing of the elliptical dataset, which includes more than 230,000 edges representing more than 200,000 transactions and money flows represented as nodes. Dataset consists of three classes: license, illegal and unknown transactions. To extract meaningful insight from this data, a broad time-based and network-based feature engineering was performed. After feature engineering, the dataset was converted into a graph structure suitable for processing by the graph neural network (GNNS). Four state -of -the -art models trained and evaluated: Graph Attention Network (GAT), GATV2, Siamese GNN, Relational Graph Convolution Network (RGCN), and graph transformers. The model was trained and validated using stratified partitions to ensure balanced class representation. In the models, the Siamese GNN achieved the highest performance with an accuracy of 95.67% GAT followed closely with 94.90% accuracy, while graph transformers 94.60% and GATV2 achieved 94.46% competitive but slightly less results. RGCN showed 93.77% the least effective performance, especially struggling with unknown transactions classification. The results strongly suggest that graphs can play an important role in increasing the accuracy and reliability of the system that detects fraud in the blockchain environment, especially in accordance with dynamic and asymmetrical graphs.