Improving Accuracy for Image Parsing Using Spatial Context and Mutual Information

Robust Method to Compute Mutual-Spatial Feature for Image Parsing Problem

“…As several approaches in image parsing area [1], [2], [5], [9], [10], [11], the labels are assigned in the region level. As mentioned in [2], this reduces computational load for the system.…”

“…Image parsing is one of challenging problems in computer vision. It is of focus in many recent works [1], [2], [3], [4], [5], [6], [7]. Firstly as noted in [2], this issue is an incredible confusing of visual words, which means one region can be matched with another region from hundreds of different labels.…”

“…Thirdly, due to the limit of the number of object labels in a parsing model, thus it is hard to build a completely plausible model, since the number of objects in the real scenes is actually unlimited. Recently, several works based on nontable parametric models for image parsing are presented in literature [1], [2], [3], [4], [5], [6], [7]. Our system is inspired from the method introduced by Tighe and Lazebnik for image parsing using scalable nonparametric model in region levels [3].…”

“…Eigen et al [5] improve MRF baseline model for image parsing problem by learning the weight of neighboring objects and Myeonget al [3] build the graph based on context model representing object relationship. Recently, Joseph Tighe et al [6] continue developing their work from [1] by learning Per-Exemplar Detectorsfeature. However, most of these works are time-consuming for learning the Manuscript received October 23, 2013; revised May 13, 2014. object relationship.…”

“…Our previous work [1] presents a novel approach for improving accuracy of the image parsing problem by using the spatial context and mutual information features in calculating the probability of object co-occurrence. The spatial context and mutual information that capture the relationship of the object location and the frequency of co-occurrence objects in one image, is a strong visual cue.…”