Research On High-resolution Saliency Detection

Deep neural network based methods have made a significant breakthrough in salient object detection.However,they are typically limited to input images with low resolutions(400 × 400 pixels or less).Little effort has been made to train neural networks to directly handle salient object segmentation in high-resolution images.This paper pushes forward high-resolution saliency detection,and contributes a new dataset,High-Resolution Salient Object Detection(HRSOD),as well as two algorithms for the proposed task.The details and contributions are described as follows:High-resolution salient object detection dataset.To facilitate studies in high-resolution saliency prediction,this paper contributes and manually annotates a high-resolution and highquality dataset,named High-Resolution Salient Object Detection(HRSOD).It contains 4010 training images and 1000 test images.For all 5010 images,pixel-level ground truths are manually annotated by over 40 subjects.To our best knowledge,HRSOD is the first high-resolution saliency detection dataset to date.Coarse-to-fine high-resolution saliency detection.Considered about blurry object boundaries in the predictions of existing saliency detection algorithms when processing highresolution images,this paper proposes a coarse-to-fine high-resolution saliency detection algorithm,which incorporates both global semantic information and local high-resolution details,to address this challenging task.More specifically,our approach consists of a Global Semantic Network(GSN),a Local Refinement Network(LRN)and a Global-Local Fusion Network(GLFN).GSN extracts the global semantic information based on down-sampled entire image.Guided by the results of GSN,LRN focuses on some local regions and progressively produces high-resolution predictions.GLFN is further proposed to enforce spatial consistency and boost performance.Salient object detection based on Deep-Shallow Residual Fusion Network and testtime fine-tuning.Existing salient object detection methods typically face two challenges.First,models are typically trained to focus more on semantics and thus the resulting segmentation boundaries are often blurry or rough.Second,due to the inductive bias on the training set,existing models do not generalize well to images with unseen contents or challenging cases.To tackle the first challenge,we propose a new model called deep-shallow network which leverages both high-level semantics and low-level boundary details to produce high-quality segmentations.Specifically,DSNet consists of a deep branch encoding semantics,a shallow branch extracting boundary details and a residual feature fusion block for consolidating advantages of both.For the second challenge,we propose a test-time fine-tuning algorithm to effectively extract imagespecific information from test images and use it to self-refine the predictions.Experiments illustrate that our proposed two methods outperform existing state-of-the-art methods on high-resolution saliency datasets by a large margin,and achieves comparable or even better performance than them on some widely used saliency benchmarks.Visual results indicate that the proposed methods are capable of accurately detecting salient objects as well as suppressing the background clutter.Further,our saliency maps have better boundary shape and are much closer to the ground truth maps in various challenging scenarios.