Jeremy S. De Bonet : Research

Image Compression
Texture Synthesis
Image Database Retrieval
Web Hacks

GIFiply: Increasing the compressibility of GIF images

In part 1 of my Ph.D. research, I've developed a image processing technique which increases the compressibility of images using Lempel-Ziv based algorithms. One algorithm of particular importance is the GIF image format which dominates data on the web today. This approach can reduce data requirements by up to 40% with virtually no visual loss. Further this approach can be combined with existing GIF reduction techniques to yield images which are as much as 60% smaller, yet still virtually indistinguishable from the originals.

Poxels: Probablistic Voxel Reconstruction

In this project I examined the problem of reconstructing a voxel representation of 3D space from a series of 2D or 1D projections of the space. We reconstruct the 3D space by optimizing over the probability that each voxel is visible in each of the projections. An iterative algorithm is used to find the optimal probability distibution which jointly explains all the observed projections.

Lempel-Ziv differential coding (LZd)

I've developed a dictionary based universal compression scheme which is able to use approximate string matches to compress data such as sounds and images in which sequences of symbols are rarely repeated exactly, but are often repeated approximately.

JPEG Blocking artifact removal

At high compression levels, the quantization of DCT coefficients used by the JPEG image format results in "blocking" artifacts. I have developed a method for reducing those artifacts by creating a new smooth basis which can recover non-blocky image from the same quantized DCT coeffients.

Texture Based Segmentation

By building up the flexible histograms texture matching technique developed in my master's thesis I have developed texture driven segmentation. Examples include segmentation of target vehicles in synthetic aperature radar (SAR), and anatomical structures from magnetic resonance imagery (MRI)

Distributions From Images

In this talk I describe a technique for modeling images by attempting to approximate such distributions directly. These approximations capture the texture characteristics within images, using an image representation which measures the joint occurrence of features across spatial resolutions. An image classification system can be designed using a similarity metric based on the likelihood that the distribution derived from one image could have generated another. Classification of natural textures indicates a high level of specificity, and recent results on target detection in SAR imagery are encouraging.

Structure Driven Image Regisration

Because of its ability to provide a representation which is generally robust to the speckle in synthetic aperatiure radar (SAR) imagery, the flexible histograms texture matching technique developed in my master's thesis can be used as core matching metric for a SAR image registration system. While working on this project during the summer of 1997 at MIT and Alphatech, Inc. I developed such a system.

Queriable Image representations

I am currently working on techniques for automatic retrieval from images databases. I developed a representation of images which Uses a complex set of filter-networks to capture the structure within a group of example images, and then use that to find similar images in a database.

Probabilistic Texture Synthesis

The textures within this web space were synthesized with a multiresolution sampling procedure which I have developed. In the first part of a two phase process, an orignal input texture is analyzed to produce a probability density estimator for the 'true' distribution from which it was generated. In the second phase, a new texture is synthesized by sampling each spatial frequency band from this density estimator. The final synthesized texture is then created by combining these spatial frequency bands.

Human Visual Psychophysics

During my final year at Columbia Engineering I became interested in studying the human visual system as a means of learning what kinds of mechanisms we employ which allow us to seemingly effortlessly make sense of the visual world around us.

Multiple Person Tracking

In the Human Computer Interface (HCI) project at the MIT Artificial Intelligence Laboratory I have developed a system to track the locations of room occupants simultaneously in 2 cameras and use that information to recover their 3D position. This information is used by many HCI room functions. It is also directly used within a tracking subsystem to direct mobile narrow-focus cameras to 'foviate' on particular occupants, and determine which of several camera views provides the most 'useful' view for recording.

Reconstructing Polyhedra From Hand-Drawn Sketches

I have built a system which can robustly reconstruct the 3D structure of rectangular polyhedra from haphazardly hand drawn sketches.

Jeremy S. De Bonet
return to main page

Page loaded on July 14, 2024 at 03:10 AM.
Page last modified on 2006-05-27
Copyright © 1997-2024, Jeremy S. De Bonet. All rights reserved.