Convolutional Neural Network Based Instance Segmentation For High-resolution Remote Sensing Images

As an important technology to extract information from high-resolution remote sensing images,instance segmentation can not only locate but also segment the individual object simultaneously.Being different from object detection which uses simple rectangles to describe objects,instance segmentation uses more refined outlines to extract more precise feature.And it has been used in the extraction of typical objects,such as buildings,vehicles,ships and so on.However,there exists several problems: one is that the orientation of objects in high resolution remote sensing images are always arbitrary,which affects the accuracy of existing instance segmentation models for the densely packed objects.Another is that labeling an instance segmentation dataset for remote sensing images is difficult and always costs a lot,and the lack of labeled data.To solve the above two problems,this thesis has carried out the related research works from the representation of target feature and semi-supervised learning based on convolutional neural network,and the main work and contributions of this thesis are as follows.For the problem of arbitrary orientation for the objects in remote sensing images,this thesis uses the rotated rectangles in instance segmentation for remote sensing images,which has improved the segmentation of arbitrary orientational and densely packed objects.Most existing instance segmentation models are based on horizontal rectangles.But for the objects in high resolution remote sensing images with arbitrary orientation or densely packed,horizontal proposals always contain other neighboring objects,and even include many background pixels,which are not conducive to get high quality features while segmenting.Thus,we proposed an instance segmentation method based on rotated rectangles,which can fit the objects well,and relieve the problems caused by horizontal rectangles.This method uses the rotated region of interest transformer to learn rotated proposals from the horizontal ones which are proposed by region proposals network.Then the rotation-invariance features extracted from rotated proposals are used for classification,location and mask segmentation.The experimental results have shown that this proposed method has obtained a good performance on the large instance segmentation dataset for high resolution remote sensing images.For the problem of high cost and difficulty to label an instance segmentation dataset and the lack of labeled data for high resolution remote sensing images,this thesis has designed a loss of mask consistency to use unlabeled images which can improve the performance of models.Though,unlabeled images have no labels,but they still contain lots of information which can be used.Based on the fact that image’s content is still consistent before and after the horizontal flip,this thesis calculates the mask consistency loss to improve the performance of models by constraining the consistency of the output,or the probability distribution of mask.The experimental results have shown that when using the same amount of labeled images,the proposed semi-supervised instance segmentation method can make better use of unlabeled images,and has obtained a higher segmentation precision than the supervised ones.