A top-down attention model based on the semi-supervised learning

In this paper, we proposed a top-down motion tracking model to detect the attention region. Many biological inspired systems have been studied and most of them are consisted by bottom-up mechanisms and top-down processes. Top-down attention is guided by task-driven information that is acquired through learning procedures. Our model improves the top-down mechanisms by using a probability map (PM). The PM follows to track if all the potential locations of targets based on the information contained in the frame sequences. By using this, PM can be regarded as a short term memory for attended saliency regions. This function is similar to the dorsal stream of V1 primary area. The semi-learning model constructs an efficient mechanism for attention detection to simulate the eye movements and fixations in our human visual systems. Generally, our work is to mimic human visual systems and it will further be applied on the robotics platform. From the random selected video clips, our performances are better than other state-of-the-art approaches.