Application of Saliency Methods for Extracting Tree Features in Outdoor Scenes

Abstract

The growing demand for accurate results in agricultural environments consolidates the so-called precision agriculture in which saliency analysis has brought possibilities for the effective application of computer vision techniques. The saliency measured by computer algorithms follows a logic of attention similar to the human visual system in which the protuberant regions are identified due to some features that make them more evident and prone to draw more attention. Thus, the salient features are preserved in such a way that the most evocative scene components are highlighted to emphasize the relevant areas. In this paper, we present a saliency map refinement approach, and we use it to compare saliency estimation methods in the detection of trees. Their performance is evaluated by counting the number of areas correctly detected and labeled as a tree, as well as the segments incorrectly categorized as a tree. We present and discuss the results to point out the salience method that best corresponds to the refinement approach we propose. Fourteen saliency methods are compared using an annotated database of manually segmented images that were collected in different scenarios where trees are emphasized in the foreground.