Dynamic Mode Decomposition for Salient Region Detection in Images

Abstract

The prime objective of this thesis is to propose versatile, efficient data-driven newlinemodels for detection of salient regions within natural images. The proposed newlinemodel should be completely unsupervised in nature (bottom up), so that it newlinecan detect any generic object, without prior learning . They should also be newlinecomputationally less complex to ensure fast processing of data. Generally, newlinecomputational models for saliency assume that the image under observation newlineis clean, and fail to account for visual disturbances like noise. The proffered newlineapproach should have the capability to deal with noise stimuli and compose newlinehigh quality saliency maps. This dissertation presents three constituent datadriven newlinesaliency models: the fist is a Dynamic Mode Decomposition (DMD)- newlinebased model, which segregates the foreground and background regions of an newlineimage as sparse and low-rank DMD modes respectively. The ensued colorand newlineluminance-based saliency maps are combined to build a full resolution newlinesaliency map. Deploying the DMD-engendered saliency map as an initial seed, newlinea Saliency-Driven Transition Region (SDTR) segmentation routine is spawned newlineto segment the salient object. To generate scalar-valued saliency maps under newlinenoisy conditions, a VMD-DMD coupled, data-driven model and a Multi- newlineResolution Dynamic Decomposition (MRDMD)-based data-driven model are newlineintroduced. The intrinsic mode functions (IMFs), obtained from 2D VMD routine, newlineare processed, and IMFS selected, based on their entropy are vectorized newlineto form a snapshot matrix, for DMD to derive low-rank and sparse components. newlineNormalized VMD modes and the DMD-kindled sparse components newlineare combined, to generate smooth saliency maps. MRDMD combines DMD newlinewith multi-resolution analysis (MRA), which models the multi-scale systems, newlinein both time and spatial domains. A sequential snapshot matrix was generated newlineusing chrominance, luminance and edge information of the image, which newlineis then given to the MRDMD module for a three-level decomposition. ..