Abstract:
Visual saliency is the perceptual quality that makes an object stand out with respect to its neighborhood and thus grabs our visual attention. Reliable estimation of visual saliency allows appropriate processing of images without prior knowledge of their contents. Hence, it proves to be an important step in many computer vision tasks such as image segmentation, image compression, object recognition, etc. Numerous visual saliency detection models have been proposed over the past two decades. These models differ in the way their features are computed and combined to generate saliency maps. Some of them are purely computational in nature and are based on mathematical and statistics background, whereas others are biologically inspired with the latest trend going towards fast yet simple implemented algorithmic models.
This work investigates major state of the art salient region detection models which includes context-aware (CA), frequency tuned (FT), global contrast using histogram-based computation (HC) and its region contrast variant (RC), and superpixel-based models. The existing superpixel-based saliency detection model uses simple linear iterative clustering (SLIC) to partition the original image into a number of superpixels.
We propose to replace SLIC with an alternative image partitioning algorithm SEEDS (superpixel extracted via energy-driven sampling) for an excellent compromise between accuracy and efficiency of superpixel-based saliency detection. We compare different salient region detection models using standard evaluation measures and datasets.
Our results show that the proposed modification in the superpixel-based saliency detection can potentially give us more than 50% improvement in terms of processing speed of model at a comparable accuracy
