Prior Guided Image Pixel-level Binary Segmentation

Image pixel-level binary segmentation task requires the network to understand the semantic information of each pixel of the input image,which could help computer perceive surrounding environment more finely and obtain useful information.In this paper,two typical pixel-level binary segmentation problems,shadow detection and RGB-D salient object detection,were used as specific tasks to study a prior guided pixel-level binary segmentation method.The shadow in the image will bring difficulties to the computer vision task and affect the correct recognition of image content.In terms of shadow detection,there are two major problems:(1)shadow detection is easy to be influenced by the background in the image;(2)it is difficult to capture accurate context of shadow area.And RGB-D saliency detection is applied in various computer vision tasks,which aims at identifying the most attractive parts from an RGB image and its corresponding depth image.However,RGB-D approaches also face two challenges:(1)how to quickly and effectively integrate the cross-modal features from the RGB-D data;(2)how to mitigate the negative impact from the low-quality depth map.We conducted the following research work for the above two tasks and their respective problems(1)We proposed a prior guided feature pyramid network for shadow detection.Firstly,we extracted the potential shadow prior information in the image and adopted the prior attention module to further filter and divide it into multi-scales prior information.Then we employed it to weigh multi-scales shadow features to help the network to locate shadow regions accurately,which increase the weight of shadow regions and reduced the weight of non-shadow regions.Secondly,we applied a feature polymerization module at the top-down pathway to fuse multi-scales shadow features step by step and used post-process operation to help the network optimize prediction results.Our proposed shadow detection method has been tested on the SBU and UCF public shadow detection benchmark datasets to evaluate its performance.Experiment results shows that the prediction results of our proposed shadow detection network are better than other previous state-of-the-arts methods in accuracy and detail(2)We introduced depth prior and proposed a depth guided residual network for RGB-D salient object detection.On the one hand,we design a simpler and efficient depth branch only using one convolutional layer and three residual modules to extract depth features instead of employing a pre-trained backbone to handle the depth data,and fuse RGB features and depth features in a multi-scale manner for refinement with top-down guidance.On the other hand,we adopted a depth correction module to evaluate the quality of depth map and add adaptive weight to depth maps to control the fusion between them,which mitigates the negative influence of unreliable depth map.Our proposed RGB-D salient object detection approach was compared with 13 state-of-the-art methods on 7 public RGB-D datasets and experimental results demonstrate the validity of the proposed approach both quantitatively and qualitatively,especially in efficiency and compactness.