A Hybrid-Backward Refinement Model for Salient Object Detection

Publication date: Available online 17 May 2019Source: NeurocomputingAuthor(s): Dakhia Abdelhafid, Tiantian Wang, Huchuan LuAbstractThe deep Convolutional Neural Networks (CNNs) have been investigated in many salient object detection works and have achieved state-of-the-art performance compared to the classic methods. However, most of the existing CNN-based methods still struggle in addressing the problem of incomplete contours of salient objects. To overcome this problem, this paper focuses on accurately capturing the fine details of salient objects by proposing a novel Hybrid-Backward Refinement Network (HBRNet), which combines the high-level and low-level features extracted from two different CNNs. Taking advantage of the access to the visual cues and semantic information of CNNs, our hybrid deep network helps in modeling the object’s context and preserving its boundaries as well. Specifically, we integrate effective hybrid refinement modules by merging feature maps of two consecutive layers from two deep networks. Also, our refinement model uses the residual convolutional unit in order to provide an effective end-to-end training. Furthermore, we apply the feature fusion technique to enable full exploitation of multi-scale features and progressively recover the resolution of the coarse prediction map. Through the experimental results, we demonstrate that the proposed framework achieves state-of-the-art performance on several popular benchmark datasets. The proposed hybrid...

Source: Neurocomputing - May 17, 2019 Category: Neuroscience Source Type: research