Attention-based Network for Image/Video Salient Object Detection

Abstract

The goal of video or image salient object detection is to identify the most important object in the scene, which can be helpful in many computer vision-based tasks. As the human vision framework has a successful capacity to effortlessly perceive locales of interest from complex scenes, salient object detection mimics a similar concept. However, the salient object detection (SOD) of complex video scenes is a challenging task. This paper mainly focuses on learning from channel and Spatiotemporal representations for image/video salient object detection. The proposed method consists of three levels, the frontend, the attention models, and the backend. While the frontend consists of VGG backbone which ultimately learns the representation of the common and the discrimination features. After that, both Attention, Channel-wise, and Spatiotemporal models are applied to highlight the significant object using a feature detector and to calculate the spatial attention. Then the output features are fused to obtain the final saliency result. Experimental investigation evaluations confirm that our proposed model has proved its validity and effectiveness compared with the state-of-the-art methods.