These days many applications of computer vision and artificial visual field in robotics process an enormous number of visual data from the charge coupled device camera. Handling this data in real-time is daunting task without the help of intelligent mechanisms. In decades, computer vision and robotics scientists have tried to effectively solve this problem on computational complexity by mimicking the concepts of human visual system or human visual perceptions. And these studies are used to build artificial visual attention model that is capable of working in real-time.
The term “saliency” is visually attentive objects or regions that appear to an observer to stand out relative to its neighbor surroundings. This kind of unique information is useful some applications such as the image segmentation, object detection, object recognition, visual tracking, robot localization, and robot navigation. Among these applications, we focused on the artificial visual attention model for image segmentation on salient regions in this paper.
There are variety of previous artificial visual attention models to detect salient regions. These models divided into the five approaches such as band-pass filter approach, domain transform approach, probabilistic approach and context-aware approach and so on. Although these models can detect locations of salient regions, they are not suitable for the image segmentation on salient regions because they have a chronic weakness point that is cannot detecting whole salient region on the given image. This disadvantage is fatal for the salient region segmentation. Moreover, their saliency map have blurring effect from multi-scale analysis, distorted shape from cluttered background. In order to overcome these disadvantages, we addressed a novel biological visual attention model that is based on the Hering’s Opponent Process Theory which describes how to perceive the visual stimuli from the human eye.
There are four main contributions. The first is a novel framework which employed the intensity and color feature channels for detecting salient regions according to the Opponent Process Theory. This framework is biologically plausible and feasible. The second is an entropy filter which detected salient regions on each feature channel. This filter can effectively and robustly detect the whole salient regions. The third is we designed the center bias map based on the central bias fixations of human beings in terms of psychological visual perception. This is used for an advance information of salient region. The last contribution is we propose the adaptive combination method on intensity and color conspicuous maps to generate final saliency map. This method has better performance than the simple linear combination method.
The experimental results in our artificial visual attention model shows proposed one is outperformed to the previous 11 artificial visual attention model on the two databases, EPFL and THUR. Specifically, the AUC(Area Under Curve) scores are 0.9708 and 0.9352 on the EPFL and THUR database, respectively. And we employed the F-measure as the evaluation method about the performance of the segmentation by the Otsu’s binarization on the detected salient regions. As a result, the precisions of proposed model is approximately higher than 88% whereas previous artificial visual attention models are lower the precision of 70%. The F-measure score of proposed model is also higher than others because of remarkable precisions. It is superior to previous models. Therefore, proposed artificial visual attention model based on the Opponent Process Theory is suitable for the image segmentation on salient regions with robustness.