Quaternion-based Region Of Interest Detection Of Color Images

With the explosive growth of image information, the issue how to remove redundant information and extract image regions of interest has been attached more and more importance. In the past 20 years, the researchers have proposed a variety of saliency detection models. Many current saliency detection approaches use semantic features to accurately capture human gaze. However, these approaches demand high computational cost and can hardly be applied to daily use. Recently, some quaternion-based saliency detection models have been proposed to meet real-time requirements, which have attracted a lot of interest. However, these models can hardly detect the object boundaries accurately and have low accuracy.In this paper, we mainly study the bottom-up models. Based on summarizing and analyzing the limitations of existing saliency detection algorithms, we put forward two saliency detection models.We propose a saliency detection model based on quarternion phase spectrum analysis and superpixel segmentation. We treat each pixel of the color image as an element of a two-dimensional quaternion matrix. The color channels will be processed as a whole in the phase spectrum extraction process to ensure that the correlation between each color channel is not destroyed. Considering that the surface of the same object should have the consistent visual saliency, we propose a local saliency map optimization strategy based on superpixel segmentation. The optimization method can effectively reduce the interference of background noise and outliers. The added motion channel allows the phase spectrum to represent spatiotemporal saliency, which means our model can detect salient regions for videos as well as images.We proposes a saliency detection model based on quaternion sparse representation residual considering the fact that regions of interest are usually difficult to be sparse represented by the neighborhood. Sparse coding decomposes the inputs into two parts, codes and residual. In this paper, we regard computing saliency values as the process of sparse coding using features of surrounding patches. The model uses center-surround structure and sparse coding to obtain saliency values for each pixel. Quaternion sparse coding can detect regions of different colors or structure at the same time while traditional sparse coding can hardly detect the regions of different colors and the same structure.We evaluate the proposed models on three commonly-used datasets using area under the ROC curve(AUC). With OSIE database, the proposed saliency detection model, which is based on quarternion phase spectrum analysis and superpixel segmentation, can reach 0.7919 area under the ROC curve value, while the quaternion-based sparse representation saliency detection model can reach 0.8054. The two models achieve better performance than most state-of-the-art saliency detection models.