Novel Mathematical/Computational Approaches to Surveillance Image Transmission and Exploitation. Phase 1
Abstract
Current image compression algorithms are founded on three basic principles: transformation, quantization, and entropy coding. The transformation attempts to remove statistical redundancies from the input, thereby reducing the data to a smaller, more manageable set. The process of quantization converts samples into a finite set of levels, typically in a way that minimizes some predefined error measure (e.g. mean-squared, mean absolute difference, etc.). The entropy coder converts the quantized output, which is in the form of a sequence of level values, into a bit string. Assignments are made with variable length codes in an attempt to minimize the overall average bit rate of the system. This study examined and developed image compression codes. Current compression methodologies such as Joint Photographers Expert Group (WEG) and others were evaluated in context of Synthetic Aperture Radar (SAR) image compression. Specifically, this benchmarking was performed in terms of visual quality and Automatic Target Recognition (ATR) performance quality. This is in direct contrast to mean-square error (MSE) and peak signal to noise ratio (PSNR) measures that are commonly used to assess image coding quality in the compression community. This important first step is done in recognition that the true measure of merit for defense applications is the preservation of target classification information after transmission. It was determined that much of the high frequency content found in SAR images was lost due to these compression techniques. Regenerating this speckle by modeling a stochastic process often reduces PSNR and MSE but increase perceptual quality. Speckle extraction and regeneration techniques were developed, implemented and evaluated on real Joint STARS SAR imagery. Image interpretation experts then analyzed and rated the compressed images.
Document Details
- Document Type
- Technical Report
- Publication Date
- Oct 26, 1999
- Accession Number
- ADA370962
Entities
People
- Charles Hsu