Dissertation

Data Dependent Optimization Vision Architecture

Lee, Chris S

2016

Overview

One of the cardinal problems in computer vision is information overload. Image sensors generate a vast quantity of pixel data through multiple channels at high speed, and such a large amount of visual data is challenging to process in real time. Although a naive approach can reduce the visual information by re-sizing images, doing so results in a data loss. To limit the data loss, recently-proposed approaches leverage \"regions of interest\", which are an intelligently selected subset of samples in an image. Unfortunately, among techniques based on generating and consuming regions of interest, each shows different performance for different scenes, which makes one fail to achieve the best performance for a set of scenes with a single technique. Motivated by this, we propose a dynamic mechanism that selects the best technique (generating the best performance) depending on the incoming scenes. In this dissertation, we propose several novel approaches to optimize computation resources by exploiting natural redundancy, dynamic algorithm selection and application specific methods. In the first part of the dissertation, we present a hardware architecture that exploits natural redundancy across pixels, frames and channels. The architecture reuses and shares the results with minimized overhead to reduce power consumption. In the second part of the dissertation, we propose a dynamic saliency algorithm selection technique that is able to choose the best saliency map based on machine learning. We demonstrate the benefits of the approach using a sampling approach across video frames. Thirdly, we show how to apply saliency maps to grocery scenes to address challenges arising for object class detection in real world scenarios. iv

• Publisher: ProQuest Dissertations & Theses
• Subject: Computer Engineering / Computer science / Electrical engineering
• ISBN: 139203969X, 9781392039694