Camera design optimization using image systems simulation
Abstract/Contents
- Abstract
- In this thesis I use physically accurate, end-to-end camera simulation environment to explore different imaging system architectures and to co-optimize the acquisition hardware with computational imaging and computer vision algorithms. The simulation tools use 3D modeling to create different scene models. Scene images that are projected onto the sensor are computed with ray tracing tools. Finally, realistic sensor and camera pipeline simulators compute pixel values that can be used by subsequent computational photography or vision algorithms. I use the simulation tools to optimize the spectral characteristics of cameras and lights used to collect data for inverse estimation methods that characterize surface spectral properties. I show that appropriately selected narrowband lights improve the performance of pixel based surface classification. I also propose a new algorithm that extends beyond narrowband light selection by computing the spectral power distribution of light that is optimal for such classification tasks. I further use simulations to guide and optimize the design of a system that characterizes fluorescent materials which interact with incident light in more complex ways than simple reflection. Spectral illuminant optimization is also useful in consumer photography. I describe a spectrally tunable flash and computational algorithms that can be used to efficiently estimate the ambient illuminant spectrum and to improve the color reproduction of captured images. Such flash is particularly useful when capturing images under extreme ambient lights, for example in underwater photography. The end-to-end nature of the simulation tools makes it possible to explore how viewing conditions (e.g. light level, spectral content, depth), hardware components of an imaging system (e.g. optics, sensor QE, filters, pixel and noise properties) and image processing pipelines affect both the perceived image quality of camera images and the ability of machine learning algorithms to detect and classify objects. I use simulation to render large collections of images in order to evaluate the object detection performance of convolutional neural networks (CNNs) . I demonstrate that the performance of CNNs trained using images that are generated by simplified image processing pipelines (ISP) is similar to the performance of CNNs trained using images generated by more complex and time-consuming ISPs optimized for consumer photography. I also explore and quantify the robustness and performance bounds of detection methods against fundamental algorithms controlling camera exposure and focus.
Description
| Type of resource | text |
|---|---|
| Form | electronic; electronic resource; remote |
| Extent | 1 online resource. |
| Publication date | 2018 |
| Issuance | monographic |
| Language | English |
Creators/Contributors
| Associated with | Blasinski, Henryk Krzysztof |
|---|---|
| Associated with | Stanford University, Department of Electrical Engineering. |
| Primary advisor | Wandell, Brian A |
| Thesis advisor | Wandell, Brian A |
| Thesis advisor | Farrell, Joyce E |
| Thesis advisor | Girod, Bernd |
| Thesis advisor | Wetzstein, Gordon |
| Advisor | Farrell, Joyce E |
| Advisor | Girod, Bernd |
| Advisor | Wetzstein, Gordon |
Subjects
| Genre | Theses |
|---|
Bibliographic information
| Statement of responsibility | Henryk Krzysztof Blasinski. |
|---|---|
| Note | Submitted to the Department of Electrical Engineering. |
| Thesis | Thesis (Ph.D.)--Stanford University, 2018. |
| Location | electronic resource |
Access conditions
- Copyright
- © 2018 by Henryk Krzysztof Blasinski
- License
- This work is licensed under a Creative Commons Attribution Non Commercial 3.0 Unported license (CC BY-NC).
Also listed in
Loading usage metrics...