Comparison of three different measures for curve saliency

Tensor voting is a saliency-based feature extraction method, which incorporates perceptual organization laws into image processing and gains its popularity in many applications, however, its saliency measure cannot adapt to some application areas, just like many bottom-up schemes measure the objective saliency of a pixel or region only based on its contrast within a local context. Here, we consider cues of the entire image in a different way. This paper puts forward two curve saliency measures for tensor voting and compares them with the original curve saliency measure in contour extraction when the density of voting tokens decreases. Experimental results show that the proposed saliency measure is more adaptive to change in voting tokens' density.