Effects of boundary artifacts in parallel implementations of wavelet compression for large-format digital framing reconnaissance camera systems

Abstract

Large-format digital reconnaissance sensors (10k x 1Ok) will soon operate at rates exceeding 100 Megapixels/sec. Operation of these airborne sensors requires lossy compression prior to real-time recording or transmission to a ground station. Wavelets have been shown to reproduce high quality imagery at both low and high compression ratios while being robust enough to efficiently compress different image classes without modifying algorithms or updating codebooks. One current implementation problem is that a full image frame is too large to compress on-board in real-time. Although parallel hardware implementations increase compression speed, undesirable reconstruction boundary artifacts can be introduced which can adversely affect the desired image analysis. This paper analyzes the effects of reconstruction artifacts on image quality.Two parallel wavelet schemes were tested, segmenting the full image frames into square Sub-Images (512x512) and into rectangular Channels (2kx128). Both schemes allow real-time operation using 8 compression processors operating at 50 segmentskec. The Sub-Image scheme equally exploits redundancy in both spatial dimensions and, for large imagery, minimizes the number or reconstruction boundaries (e.g. 38 for a lOkxlOk image compared to 83 for the Channel scheme). Twelve digital airbome reconnaissance images taken at different altitudes (5-25kft) with varying scene content and image complexity were used in the simulations to ensure robustness of the results.The table below shows combined objective performance results using a 40:l compression ratio, which would allow transmission of 1-2 framedsec over a single 42.8 Mbps Common Data Link sub-channel. The reconstruction boundary edges were visible using both parallel schemes and, although the perceivable ground resolvable distance (measured by MIL-STD 150A Photo Resolution Bar Target analysis) was reduced by the 40: 1 wavelet compression itself, no such resolution loss was caused by the reconstruction edges. Relative to compressing the full 10k x 10k image, the parallel schemes introduce minimal objective or subjective image degradation and reduce the required compression timc more than 8x.