We propose an improved saliency guided wavelet
compression scheme for low-bitrate image/video coding applications.
Important regions (faces in security camera feeds,
vehicles in traffic surveillance) get degraded significantly at low
bitrates by existing compression standards, such as JPEG/JPEG-
2000/MPEG-4, since these do not explicitly utilize any knowledge
of which regions are salient. We design a compression algorithm
which, given an image/video and a saliency value for each
pixel, computes a corresponding saliency value in the wavelet
transform domain. Our algorithm ensures wavelet coefficients
representing salient regions have a high saliency value. The
coefficients are transmitted in decreasing order of their saliency.
This allows important regions in the image/video to have high
fidelity even at very low bitrates. Further, our compression
scheme can handle several salient regions with different relative
importance. We compare the performance of our method with
the JPEG/JPEG-2000 image standards and the MPEG-4 video
standard through two experiments: face detection and vehicle
tracking. We show improved detection rates and quality of reconstructed
images/videos using our Saliency Based Compression
(SBC) algorithm.