Date of Award
Doctor of Philosophy (PhD)
Dr Michael Weeks
Dr Rajshekhar Sunderraman
Dr Juan Banda
Dr George Pullman
In recent times, digital media contents are inherently of multimedia type, consisting of the form text, audio, image and video. Several of the outstanding computer Vision (CV) problems are being successfully solved with the help of modern Machine Learning (ML) techniques. Plenty of research work has already been carried out in the field of Automatic Image Annotation (AIA), Image Captioning and Video Tagging. Video Captioning, i.e., automatic description generation from digital video, however, is a different and complex problem altogether. This study compares various existing video captioning approaches available today and attempts their classification and analysis based on different parameters, viz., type of captioning methods (generation/retrieval), type of learning models employed, the desired output description length generated, etc. This dissertation also attempts to critically analyze the existing benchmark datasets used in various video captioning models and the evaluation metrics for assessing the final quality of the resultant video descriptions generated. A detailed study of important existing models, highlighting their comparative advantages as well as disadvantages are also included.
In this study a novel approach for video captioning on the Microsoft Video Description (MSVD) dataset and Microsoft Video-to-Text (MSR-VTT) dataset is proposed using supervised learning techniques to train a deep combinational framework, for achieving better quality video captioning via predicting semantic tags. We develop simple shallow CNN (2D and 3D) as feature extractors, Deep Neural Networks (DNNs and Bidirectional LSTMs (BiLSTMs) as tag prediction models and Recurrent Neural Networks (RNNs) (LSTM) model as the language model. The aim of the work was to provide an alternative narrative to generating captions from videos via semantic tag predictions and deploy simpler shallower deep model architectures with lower memory requirements as solution so that it is not very memory extensive and the developed models prove to be stable and viable options when the scale of the data is increased.
This study also successfully employed deep architectures like the Convolutional Neural Network (CNN) for speeding up automation process of hand gesture recognition and classification of the sign languages of the Indian classical dance form, ‘Bharatnatyam’. This hand gesture classification is primarily aimed at 1) building a novel dataset of 2D single hand gestures belonging to 27 classes that were collected from (i) Google search engine (Google images), (ii) YouTube videos (dynamic and with background considered) and (iii) professional artists under staged environment constraints (plain backgrounds). 2) exploring the effectiveness of CNNs for identifying and classifying the single hand gestures by optimizing the hyperparameters, and 3) evaluating the impacts of transfer learning and double transfer learning, which is a novel concept explored for achieving higher classification accuracy.
Perunninakulath Parameshwaran, Anuja, "Deep Architectures for Visual Recognition and Description." Dissertation, Georgia State University, 2020.
File Upload Confirmation