Deep Learning Based Visual Saliency Detection

Visual saliency refers to that there is always a local area in the visual field that can arouse the most attention of the visual system when people observe a certain area,which is called the saliency area.Visual saliency detection is primarily used to highlight the saliency area in the image or video.With the rapid development of computer vision and the wave of artificial intelligence,using computer technology to simulate the human eye's attention mechanism has become an emerging and challenging research hotspot.At the moment,visual saliency detection is widely used in image segmentation,object detection,video coding and other fields.Therefore,it is of great significance to carry out the research of visual saliency detection.At present,visual saliency detection is roughly divided into RGB image saliency detection,RGBD image saliency detection,video saliency detection and Co-saliency detection.In this paper,deep learning technology is introduced to study the saliency detection of RGBD image and video.Firstly,the deep learning algorithm is used to fully exploit the complementary information between RGB image and depth image,and the saliency detection of RGBD is effectively realized by multi-scale feature fusion strategy.A method of RGBD image saliency detection based on multi-scale feature fusion is proposed.Secondly,aiming at the characteristics of video saliency detection task,this paper ingeniously introduces attention mechanism,feature hierarchical fusion strategy and boundary information fusion strategy,and proposes a deep fusion video saliency detection model which introduces attention mechanism.In this paper,we use deep learning algorithm and introduce multi-scale strategy and attention mechanism to design the corresponding RGBD image saliency detection model and video saliency detection model to realize the detection of saliency regions or objects,which effectively promote the development of visual saliency detection.Through a large number of comparative experiments,the proposed models have a large performance improvement on multiple datasets.