Dual-path Processing Network for High-resolution Salient Object Detection

In the trend of deep learning, salient object detection (SOD) has achieved preeminent performance over the years. Nevertheless, scant work has been done to address the contradiction between excellent saliency map segmentation and the computational consumption required by high-resolution input images (e.g., 1024x2048 pixels or more). To meet this challenge, we introduce a dual-path processing network (DPPNet) to detect and segment salient objects in high-resolution input images directly and efficiently, the network contains global context and spatial details paths. Specifically, the global context path utilizes a multi feature extraction and enhanced (MFEE) module to extract richer global multiscale semantic features with a large receptive field at a lower resolution. The spatial details path employs a boundary information guided (BIG) module to focus on accurate saliency objects location and maintain local boundary information at a higher resolution. Guided by the BIG module, a feature fusion unit (FFU) is further employed to heighten the spatial consistency of maps at different levels and boost the robustness of the network. Extensive evaluations on two high-resolution SOD datasets and four low-resolution SOD mainstream datasets indicate that the method we proposed can settle the challenge of high-resolution image input effectively and exceed ten top-notch comparison algorithms.