Saliency Driven Imagery Preprocessing for Efficient Compression -- Industrial Paper

TL;DR

The paper addresses the growing storage and bandwidth challenges of satellite imagery by introducing a saliency-driven preprocessing step that applies per-pixel, variable-width Gaussian smoothing guided by saliency masks before traditional compression. This approach yields spatially adaptive information loss, preserving salient regions while achieving storage savings when paired with JPEG2000 or BPG. Across multiple remote-sensing datasets, the method demonstrates favorable trade-offs between rate reduction and reconstruction error, and downstream object detection can be preserved or even improved at equivalent bit-rates. The work has practical implications for edge storage and transmission in remote sensing, enabling more efficient data pipelines without redesigning end-to-end codecs.

Abstract

The compression of satellite imagery remains an important research area as hundreds of terabytes of images are collected every day, which drives up storage and bandwidth costs. Although progress has been made in increasing the resolution of these satellite images, many downstream tasks are only interested in small regions of any given image. These areas of interest vary by task but, once known, can be used to optimize how information within the image is encoded. Whereas standard image encoding methods, even those optimized for remote sensing, work on the whole image equally, there are emerging methods that can be guided by saliency maps to focus on important areas. In this work we show how imagery preprocessing techniques driven by saliency maps can be used with traditional lossy compression coding standards to create variable rate image compression within a single large satellite image. Specifically, we use variable sized smoothing kernels that map to different quantized saliency levels to process imagery pixels in order to optimize downstream compression and encoding schemes.

Figures (17)

• Figure 1: Rate (bpp) distribution across selected datasets.
• Figure 2: Zoomed in crops of various smoothing kernels applied to the whole image.
• Figure 3: Generated Saliency Masks
• Figure 4: Examples of various saliency-masks applied with spatial-blurring.
• Figure 5: Comparison of spatially smoothed composites to their uniformly smoothed counterparts. Since spatially smoothed composites target low saliency regions for higher compression, we show effective trade-off of information encoding between low and high imporatance regions