Deep learning has achieved great success in recent years in computer vision and its related areas. For core computer vision tasks such as image classification, image semantic segmentation, image super-resolution, and object detection from images, deep learning based methods outperform various traditional methods in terms of both accuracy and speed. While a myriad of deep learning based computer vision research projects are continuously pushing forward the frontier of computer vision further by improving the performance for image-level tasks, many recent investigations have begun to look into deep learning based methods for sequential data such as videos and medical image sequences.With the extra information from its additional sequential dimension, sequential data naturally raises an important and challenging question: How can we effectively and efficiently integrate such sequential information into existing successful and sophisticated image-based deep learning frameworks without building from scratch?In this dissertation we develop techniques and methods that enable us to incorporate sequential information into existing image-based deep learning frameworks for different computer vision tasks. More specifically, we propose advanced methods that successfully utilize both image-based deep learning models and sequential information for the super-resolution task using multi-slice computed tomography image sequences, and for the object detection and tracking task using multi-frame videos. We demonstrate how we integrate sequential information into modern image-based deep learning systems for these different tasks under different integration paradigms. Our experiments show that our proposed methods have significantly improved the performances compared with naive image-based methods, and achieved the new state-of-the-art for such sequential vision tasks.
PDF
download
878987797
028-85220240
