Research On Salient Object Detection And Camouflage Instance Segmentation Methods

With the rapid improvement of hardware performance,the popularity of smart mobile devices,and the increasing speed of the network,images have become a mainstream channel for people to obtain information and communicate with each other.At the same time,the information in the image also increases,how to efficiently extract the region of interest from the image becomes more important for scene understanding.Salient object detection and camouflage detection study this project from two different aspects.At present,most salient object detection algorithms capture salient objects by solely extracting the image feature,but how to make full use of other auxiliary information remains to be studied.However,it is more difficult to distinguish the camouflaged objects in the scene because the camouflaged objects share similar color and texture characteristics with the background.On this basis,it is even more challenging to distinguish different camouflaged instances.In this paper,we will conduct research based on these two problems.In the research on salient object detection,we propose a general dynamic salient object detection model,called DANet,which can integrate arbitrary auxiliary information to help detect salient objects without changing the main network.We conducted experiments to study the roles of different auxiliary information and demonstrated the universality of DANet in processing auxiliary information.DANet obtained max-F,max-E,and MAE of0.87,0.9,and 0.04 respectively on the DUTS-TE dataset,which are better than the mainstream models.The salient object detection model extracts the salient object of the scene,while the camouflage detection model captures the object of the camouflaged object from the background.Although they are both category-independent pixel-level segmentation tasks,there are significant differences between their features,so the existing salient object detection models have limitations in extracting camouflaged objects.In the research on camouflaged instance segmentation,we focus on extracting global information and propose an instancelevel camouflaged object detection model CISTR based on Transformer,which can extract camouflaged objects and distinguish different instances.CISTR combines convolutional neural network and Transformer to model both local and remote connections.This Transformer based method achieves 38.6% AP on instance labels of COD10 K test set,which exceeds other mainstream instance segmentation models.