PROJECT TITLE :

PISA: Pixel wise Image Saliency by Aggregating Complementary Appearance Contrast Measures With Edge-Preserving Coherence - 2015

ABSTRACT:

Driven by recent vision and graphics applications like image segmentation and object recognition, computing pixel-correct saliency values to uniformly highlight foreground objects becomes increasingly necessary. In this paper, we have a tendency to propose a unified framework referred to as pixelwise image saliency aggregating (PISA) numerous bottom-up cues and priors. It generates spatially coherent yet detail-preserving, pixel-accurate, and fine-grained saliency, and overcomes the limitations of previous ways, which use homogeneous superpixel based and color solely treatment. PISA aggregates multiple saliency cues during a world context, such as complementary color and structure contrast measures, with their spatial priors within the image domain. The saliency confidence is further jointly modeled with a region consistence constraint into an energy minimization formulation, in that each pixel will be evaluated with multiple hypothetical saliency levels. Rather than using international discrete optimization methods, we tend to employ the price-volume filtering technique to unravel our formulation, assigning the saliency levels smoothly while preserving the sting-aware structure details. Still, a faster version of PISA is developed using a gradient-driven image subsampling strategy to greatly improve the runtime efficiency whereas keeping comparable detection accuracy. In depth experiments on a number of public information sets recommend that PISA convincingly outperforms other state-of-the-art approaches. Also, with this work, we tend to conjointly produce a brand new knowledge set containing 80zero commodity images for evaluating saliency detection.