Salient object detection using learning classifier systems that compute action mappings

Learning classifier systems (LCSs) are rule-based online evolutionary machine learning techniques that solve a problem by interacting with an environment. LCSs have been successfully used in various applications such as data mining, robot control and computer vision systems. Salient object detection is the task of automatically localizing the objects of interests in a scene by suppressing the background information, which facilitates various machine learning applications such as object segmentation, recognition and tracking. It is a difficult problem as natural scenes can often have objects with cluttered backgrounds (making it difficult to distinguish the object from background based on its features) or other complicating factors such as multiple objects. Existing saliency learning methods learn a single weight vector emphasizing the importance of each feature/attribute for the whole image dataset, hence losing generalization in the test phase when considering unseen images. LCS technique has the ability to learn weight sets for different types of images automatically. Hence, this paper investigates the application of LCS for learning image dependent feature fusion strategies for the task of salient object detection. Our LCS approach evolves generalized rules for a well known benchmark dataset consisting of 1000 images, of various types and difficulty levels, and outperforms a genetic algorithm based system that was previously state-of-the-art.