Importance coding of surveillance imagery for interpretability using quadtree dynamic importance maps

Abstract

This paper presents an importance coding technique to improve the interpretability versus bit-rate performance for an image compression system. The term interpretability is a subjective image quality measure of the content recognition performance by trained image analysts. A quadtree dynamics importance map is used to aid the prioritisation of the image encoded bit-stream according to its importance for interpretability. The importance map is generated using several bottom-up, context-free features for the identification of regions-of-interest in surveillance imagery that correlate well with visual attention processes in humans. Traditional PSNR progressive coders rely on descending wavelet coefficient magnitudes to prioritise the encoded bit-stream. The importance coder discussed in the paper is designed such that the JPEG2000 standard can implement the importance prioritisation schema (without the need to send the importance map information). Subjective evaluations indicate that the proposed schema is better than traditional PSNR progressive coders for the purpose of very low bitrate content recognition.