Research On Prior Knowledge Based Image Segmentation Technology

Image segmentation is the basics of understanding image content.An increasing number of applications nourish from inferring knowledge through it.Some of those applications include autonomous driving and computational photography to name a few.Although the fully convolution methods based on deep learning have made great progress,the model of pixel-level multi-classification relies on a lot of expensive annotation works.However,due to the lack of annotation data in practical application,segmentation results are often limited by training data,and the robustness is not good in changing scenarios.In view of the above problems,this thesis aims to study a more robust image segmentation method,explores how to introduce priori information to guide and promote the learning of the network,so as to train the model by using the pixel annotation with only fewer categories and fewer images,and realize the effective segmentation of categories and changing scenes without pixel annotation.It mainly includes the following two aspects:(1)In order to solve the problem of lack of pixel annotation in semantic segmentation and instance segmentation,this thesis proposes a segmentation model based on the location priori to make mask prediction for each object and explore the shared segmentation features among different objects.For partial supervised semantic segmentation,the bounding box annotation can be effectively converted to pixel level strong annotation.For partial supervised instance segmentation,only a few categories of pixel annotation are used to realize instance segmentation of object categories without pixel annotation in the training set,in combination with the object detection algorithm.Specifically,when given an image,through quite mature automatic detection methods or human interactions,it is easy to get the bounding box as a guiding input of the network.The Gradient Gaussian Attention Generator creates localizable and discriminative attention maps based on the bounding box to facilitate forward propagation.The multi scale alignment loss function considers local and global clues together and strengthens the constraint to facilitate back propagation.Multiple experiments were carried out on PASCAL VOC,MS COCO and other datasets.Compared with the existing methods,the segmentation accuracy has been significantly improved.(2)In order to improve the robustness of ground segmentation,a robust two-branch ground segmentation algorithm based on geometric priori is proposed.Through data analysis and statistics,disparity images and ground plane deviation maps obtained by fitting the ground plane have stable distribution on different datasets and contain specific plane geometric properties.This thesis introduces the geometric prior to promote the forward propagation of the network.The main branch of the network adopts convolution module with multiple sampling rates to extract the multi-level structure of features and effectively smooth the empty information in geometric priori by combining local and global context information.The enhancement module adaptively generates the feature maps of the reference geometric structure,which improves the robustness of the segmentation results.The ground segmentation model is evaluated under six different scenarios,and the segmentation results are more robust than the existing methods.