|
Surveillance imaging applications on small autonomous imaging platforms present challenges of highly constrained power supply and form factor, with potentially demanding specifications for target detection and recognition. Absent of significant advances in image processing hardware, such power and space restrictions can imply severely limited computational capabilities. This holds especially for compute-intensive algorithms with high-precision fixed- or floating- point operations in deep pipelines that process large data streams. Such algorithms tend not to be amenable to small or simplified architectures involving (for example) reduced precision, reconfigurable logic, low-power gates, or energy recycling schemes. In this series of two papers, a technique of reduced-power computing called compressive processing (CXP) is presented and applied to several low- and mid-level computer vision operations. CXP computes over compressed data without resorting to intermediate decompression steps. As a result of fewer data due to compression, fewer operations are required by CXP than are required by computing over the corresponding uncompressed image. In several cases, CXP techniques yield speedups on the order of the compression ratio. Where lossy high-compression transforms are employed, it is often possible to use approximations to derive CXP operations to yield increased computational efficiency via a simplified mix of operations. The reduced work requirement, which follows directly from the presence of fewer data, also implies a reduced power requirement, especially if simpler operations are involved in compressive versus noncompressive operations. Several image processing algorithms (edge detection, morphological operations, and component labeling) are analyzed in the context of three compression transforms: vector quantization (VQ), visual pattern image coding (VPIC), and EBLAST. The latter is a lossy high-compression transformation developed for underwater communication of image data. Theory is primarily based on our previous research in compressive target detection and recognition. The modeling technique that supports analysis and verification of claims emphasizes 1) translation of each algorithm to a given compressive format, 2) determination of the operation mix M for each algorithm produced in Step 1), and 3) simulation of M on various architectural models, to estimate performance. Where possible, algorithms are expressed in image algebra, a rigorous, concise notation that unifies linear and nonlinear mathematics in the image domain. Image algebra has been implemented on a wide variety of workstations and parallel processors, thus increasing portability of the algorithms presented herein.
|
