Fast image adaptation driven by people presence

This paper describes an image adaptation system based on the management of regions of interest (ROIs) before its coding. The system provides a set of tools for ROI classification, filtering and merging in an efficient way. The proposed tools can be applied to any binary segmentation mask created by any semantic content-based algorithm, but in this paper the system is customized to drive the adaptation based on people presence. In order to detect presence of people, the segmentation algorithm works looking for flesh colour zones. In order to get a final set of regions of interest (ROIs), after the flesh colour regions are found they are filtered and merged depending on their size, shape, density and other parameters values customized by the conditions imposed by looking for people. Once we have the final set of ROIs there are two possibilities for coding. The first possibility is to code them using the JPEG2000 ROI encoding capabilities which provide file size reduction by codifying the less important zones of the image with reduced quality. The other possibility is the selection of main zones of the image fitting in a particular spatial resolution (aimed to limited resolution terminal adaptation) trying to englobe the most significant ROIs.