An efficient, SELective, Perceptual-based super-resolution estimator

Efficient Super-Resolution driven by saliency selectivity

“…At each iteration of the SR estimation, all the pixels are processed on an HR grid inclusively, thus, demanding high computational requirements in solving the inverse problem, which makes these methods impractical in real-time environments or under platforms of low computational specifications. Consequently, in order to minimize the computational complexity, the selective SR framework as presented in [7][8][9], processes only a subset of active pixels at each iteration of the SR solution. In the selective SR framework, the gradient descent iterative solution [2][3][4] of the SR problem (2) is modified as follows:…”

“…In [7], a selective perceptual mask is presented such that active pixels are detected by exploiting the contrast sensitivity of the Human Visual System (HVS). Moreover, in [8,9], the high-complexity saliency-based visual attention model of [10] is adopted to further reduce the set of active pixels.…”

“…The contrast sensitivity threshold is the measure of the smallest contrast, or Just Noticeable Difference (JND), that yields a visible signal over a uniform background [7]. Ferzli et al [7] proposed a model where the contrast sensitivity thresholds are computed in the spatial domain using a sliding window block of size .…”

“…Towards this issue, a selective approach of SR estimators [6][7][8][9] is introduced to reduce the dimensionality or the computational complexity of popular SR algorithms without a noticeable degradation in the desired HR reconstruction quality. These selective algorithms detect only a subset of active pixels to be processed in an iterative SR solution based on a local significance measure.…”

“…This technique is impractical due to manually tweaking the gradient thresholds for each image differently in order to attain the best desired SR quality. In [7], a set of significant pixels is determined adaptively using an automated perceptual detection mechanism without any manual tuning. Due to Human Visual Attention, attended regions are processed at high visual acuity, hence details present in these regions should be reconstructed with higher accuracy than those present in non-attended areas.…”