First Advisor

Y. C. Jenq

Term of Graduation

Spring 1999

Date of Publication

5-6-1999

Document Type

Thesis

Degree Name

Master of Science (M.S.) in Electrical and Computer Engineering

Department

Electrical and Computer Engineering

Language

English

Subjects

Charge coupled devices, Night vision, Adaptive filters, Neural networks (Computer science)

Physical Description

1 online resource (102 pages)

Abstract

Sensor saturation in the presence of bright objects can create a non-restorable loss of image information as the spatial gradient tends to zero. Blooming is an attendant morphological corruption of image data that plagues quantum well sensors, and in particular, low-light Charge-Coupled Devices (CCD's) used for nightvision applications. A novel application of Radial Basis Function Neural Networks has shown promise as a means for patching gain/brightness adjusted short exposure frame detail into the bloomed regions of long exposure image frames. A computer-tethered CCD camera was used to capture robust target images, at different exposures, in a controlled light-box environment to enable experimental image fusion. Histogram specification was demonstrated to be an effective means of mapping the gain and brightness of the shorter exposure, patch frame to the longer exposure, host frame. Alpha channel mixing could then provide a structured and easily controlled means of creating composite scenes in real time.

The alpha mask that identifies bloomed regions was generated either by trained RBFNN's, or by conventional threshold techniques. Determination of alpha was based on the filtered difference between long exposure frames and gain/brightness adjusted short exposure frames. Results showed that even conventional means of segmentation using simple thresholds produce improved composite images for nightvision applications, by extending global dynamic range and reducing distracting blooming glare. RBFNN's combined with multiscale filtered difference inputs demonstrated adaptive behavior to balance tradeoffs between fusion smoothness and image detail preservation near bloomed regions, based on training target complexity.

Traditional image restoration techniques, such as Wiener or Kalman filtering, are based on inverting the effects of known or postulated image degradation models, and are not defined for dyadic image operations. Neural networks have an advantage achieving image combination by learning from example. Neural networks also have the advantage of physiological parallels to the local receptor processing of the human visual system, and can emulate perceptive processes for synthetic vision.

Rights

In Copyright. URI: http://rightsstatements.org/vocab/InC/1.0/

This Item is protected by copyright and/or related rights. You are free to use this Item in any way that is permitted by the copyright and related rights legislation that applies to your use. For other uses you need to obtain permission from the rights-holder(s).

Persistent Identifier

https://archives.pdx.edu/ds/psu/45134

Download

Share

COinS