Proceedings Volume Mathematics of Data/Image Coding, Compression, and Encryption with Applications XII, 77990K (2010) https://doi.org/10.1117/12.864279
We live in such perceptually rich natural and manmade environments that detection and recognition of objects is mediated cerebrally by attentional filtering, in order to separate objects of interest from background clutter. In computer models of the human visual system, attentional filtering is often restricted to early processing, where areas of interest (AOIs) are delineated around anomalies of interest, then the pixels within each AOI's subtense are isolated for later processing. In contrast, the human visual system concurrently detects many targets at multiple levels (e.g., retinal center-surround filters, ganglion layer feature detectors, post-retinal spatial filtering, and cortical detection / filtering of features and objects, to name but a few processes). Intracranial attentional filtering appears to play multiple roles, including clutter filtration at all levels of processing - thus, we process individual retinal cell responses, early filtering response, and so forth, on up to the filtering of objects at high levels of semantic complexity.
Computationally, image compression techniques have progressed from emphasizing pixels, to considering regions of pixels as foci of computational interest. In more recent research, object-based compression has been investigated with varying rate-distortion performance and computational efficiency. Codecs have been developed for a wide variety of applications, although the majority of compression and decompression transforms continue to concentrate on region- and pixel-based processing, in part because of computational convenience. It is interesting to note that a growing body of research has emphasized the detection and representation of small features in relationship to their surrounding environment, which has occasionally been called semantic compression.
In this paper, we overview different types of semantic compression approaches, with particular interest in high-level compression algorithms. Various algorithms and approaches are considered, ranging from low-level semantic compression for text and database compaction, to high-level semantic analysis of images or video in which objects of interest have been detected, segmented, and represented compactly to facilitate indexing. In particular, we overview previous work in semantic pattern recognition, and how this has been applied to object-based compression. Discussion centers on lossless versus lossy transformations, quality of service in lossy compression, and computational efficiency.