On the convergence of image compression and object recognition

Abstract

Over the past four decades, image compression and object recognition have evolved from pixel-level to region-level processing, and thence to feature-based resolution. For example, compression has progressed from entropy coding of a bit or pixel stream, to transform coding applied to rectangular encoding blocks, to feature-based compression that employs segmentation of isospectral or isotextural regions. Similarly, object recognition has progressed from operations on individual pixel intensities or spectral signatures, to region-or feature-based processing employing segmentation. Recent progress in image compression indicates that significantly decreased bit rate (thus, significantly increased compression ratio) can be obtained by isolating or segmenting, then compressing scene objects, which is called object-based compression or OBC.In this paper, the relationship between object segmentation and compression is presented theoretically, then exemplified in terms of current research in OBC. The relationship between object compression and recognition is also discussed theoretically. Recent work in object recognition is shown to be closely related theoretically to OBC. Two recently-developed paradigms, Muñoz et al.'s region growing algorithm and Campbell et al.'s QuickSketch, which support very efficient compression and accurate recognition/retrieval of image objects, are summarized. Additionally, a performance model is given for OBC that isolates space complexity bottlenecks for future enhancement and implementation.

<title>Error measures for object-based image compression</title>

“…Section 2.1 overviews image algebra notation employed in this paper, while Section 2.2 presents a summary of OBC. Section 2.3 summaries two published segmentation algorithms that are the core process in the first author's published implementation of OBC [14], and Section 2.4 discusses their relationship to the semantic description of images.…”