Research On Salient Object Detection Based On Hard Sample Mining

As the most commonly used information carrier in human society,there is often a large amount of information in images.The human's visual perception system can quickly locate the most attention-grabbing content in complex images.Researchers hope that computers can mimic it and have the ability to automatically locate salient content,and so the task of salient object detection was proposed.The task aims to automatically identify and locate the most attractive regions from the input image.As a basic and effective preprocessing method,it not only helps to reduce the complexity of scene analysis,but also is widely used in image and video compression,semantic segmentation,visual tracking and other fields with high practical value.With the development of deep convolutional neural networks and the emergence of numerous salient object detection datasets,significant progress has been made in this field.However,in the face of more complex scenes,most saliency methods still fail to accurately identify salient objects,which usually manifests itself in the presence of some pixel-level hard samples with low confidence in the prediction.To address the above problems,this paper focuses on hard sample mining and feature fusion for salient object detection.The main work of this paper is as follows:First,the poor detection in complex scenes is due to the fact that most current methods lack attention to pixel-level hard samples in images,which makes it difficult for models to extract features with sufficient discrimination at corresponding locations.To address this problem,a progressive salient object detection method based on weak feature is proposed to try to refine the saliency map by enhancing the features corresponding to the hard samples.The method extracts hard-to-predict regions(low-confidence regions)in the image,and target-enhances the features in these regions by a carefully designed weak feature boosting module.Starting from a coarse saliency map,it is gradually refined based on the enhanced features,resulting in the more accurate saliency map.The evaluation results on five benchmark datasets show that the saliency map using this method is completer and more accurate,which has advantages over the state-of-the-art methods.Second,to address the problem of dramatic changes in the scale of salient objects,the dynamic scale-aware learning method is proposed by introducing the dynamic routing mechanism into the salient object detection task,which includes dynamic fusion methods at both intra-feature and inter-feature levels to achieve learning adaptive routing methods for salient objects at different scales.Then,the sampling and refinement strategy is proposed for the hard sample problem.The method generates probabilistic maps of different classes under the guidance of the coarse saliency map,selects pixels by random sampling and introduces the graph attention mechanism to construct appropriate graph representations to finally achieve targeted semantic information transfer and feature aggregation for hard samples.The experimental results on the above datasets show that the saliency maps generated by this method can better cope with the scale and the hard sample problem,and also show excellent performance in both quantitative and qualitative aspects.