Visual Saliency Based Adaptive Image Compression in JPEG2000

Abstract

While compressing the images, their purpose must be considered. The natural and advertisement images contain salient region(s) that grab viewer s attention. In other words, the common or non-salient parts are just ignored. Thus, it is necessary to reconstruct such images with the best newlinepossible quality in the salient parts or Region of Interest (ROI) and non-noticeable artefacts in the background. It will hold viewers attention in the areas of higher quality that may ignore some coding artefacts occurring in the background. newlineThis research work explores the concept of visual saliency and its application in Region newlineof Interest (ROI) prioritized image coding. Investigation and development of computationally efficient models for visual saliency estimation in images is one of the objective. Further, new methods that support both multiple ROI coding and fine delimiting of ROI are proposed and implemented in the framework of JPEG2000. Finally, the integration of proposed fast saliency detection and multiple ROI coding methods using an adaptive k-mean clustering is demonstrated for natural and advertisement image compression. The introduced spectrum scale space reduction modules speed up the saliency detection newlineof Hypercomplex Fourier Transform (HFT) model, at least two times, with the same saliency newlineestimate. The developed saliency-based multiple ROI coding algorithm changes rate-distortion newlinefactor according to the saliency in ROIs. It utilizes the rate distortion factor to drive the postcompression rate-distortion optimization in JPEG2000 to rearrange the visually salient regions in the bit-stream preferentially. These methods maintain the visual quality in most salient part and it gradually drops towards the background as per the saliency in different regions. Maxshift and Implicit are the two JPEG2000 compliant ROI coding methods.