Research On High-resolution Salient Object Detection

Salient object detection aims to accurately detect and segment the most representative object region in the image,which is often applied to other computer vision tasks as a pre-processing step.Due to hardware limitations,previous saliency detection algorithms are aimed at low-resolution image input,and scant studies on saliency detection directly for high-resolution images（e.g.,1024×2048 pixels）.With the widespread use of high-definition electronic devices,image quality is mostly high-resolution and has increased the performance requirements for image processing.In view of this,this paper carries out the research of high-resolution salient object detection algorithm under fully supervised and weakly supervised learning.High-resolution Salient Object Detection under Fully Supervised:This paper presents a dual-path processing network for high-resolution salient object detection under fully supervised learning.The network contains both the global context path and the spatial detail path.The global context path utilizes a revised backbone and a multiscale feature enhancement module to gradually extract rich global multiscale semantic features with large receptive field.The spatial detail path mainly focuses on the accurate saliency boundary information and uses the boundary information to guide the module to retain salient object boundary details.Finally,the network improves the spatial consistency of feature maps at different levels through feature fusion units and obtains salient maps with clear boundaries.Extensive experiments on two high-resolution and four low-resolution salient object detection datasets show that the proposed algorithm under fully supervised learning achieves a balance between speed and accuracy for high-resolution image processing,with F^maxvalues of 0.906 on the HRSOD dataset and 0.957 on the ECSSD,and it can also obtain better prediction results of clear edges in complex scenes.High-resolution Salient Object Detection under Weakly Supervised:The fully supervised learning model relies on large-scale dataset training with per-pixel annotation and is limited by the number of dataset samples.This paper further develops weakly supervised learning-based high-resolution salient object detection for scribbles.A scribble annotated high-resolution salient object detection dataset,Scr-HRSOD,is first created,and then a two-branch network incorporating different resolution inputs is proposed to iteratively mine correlations between adjacent pixels in combination with scribble labelled partially labelled pixels to progressively propagate partial pixel information to unknown regions,while obtaining the object boundary constrained by the spatial location attention module capturing the complete salient object segmentation.Experimental results on scribble dataset and five other datasets show that the weakly supervised learning algorithm proposed in this paper achieves higher values of 0.849 F^max and 0.867 S_m on HRSOD dataset,which is superior to the weakly supervised and unsupervised algorithms in contrast algorithm,and even exceeds the partially fully supervised algorithm.The visualization results show that the algorithm based on graffiti supervision can accurately detect and segment salient objects in different scenes.