UC Irvine

The field of computer vision has seen tremendous growth in recent years. However, traditional computer vision models, machine learning algorithms, and image processing techniques rely on an immense number of parameters and require large datasets of annotated data. This is where convolutional neural networks present a viable solution. These networks provide a more efficient approach by requiring fewer parameters compared to other neural networks, while still delivering remarkable feature extraction abilities. Despite their widespread usage, primarily as a backend for feature extraction, a significant number of applications do not fully exploit the potential of convolutional layers and are prone to issues such as vanishing or small gradients, especially when employed as encoders in the initial layers. To fully leverage the capabilities of CNNs, it is crucial to understand their functioning and employ them in a manner that harnesses their full potential. Therefore, deep fully convolutional neural networks have become a popular choice for various tasks, particularly in the semantic segmentation of images. Semantic segmentation is characterized by its dense prediction capability, achieved through the assignment of a class label to every individual pixel. In this dissertation, we delve into the realm of deep fully convolutional architectures for semantic segmentation. Following an introduction to the basic blocks of deep neural networks, Chapter two presents a review of well-known models and evaluation metrics for semantic segmentation. Chapters three and four explore pre-processing and various tasks applied to satellite images and a review of deep-learning models for financial time-series prediction. The review highlights the current gap in the application of fully convolutional networks for financial time-series analysis and classification. Chapter five, with an aim to address limitations in current designs for small datasets with high-resolution images, proposes a modified U-Net to detect war-inflicted damage in satellite images. Finally, shifting toward time-series data, chapter six establishes an accurate and efficient fully convolutional encoder-decoder architecture for multi-variate stock price trend forecasting, utilizing daily prices time-series.