Research On High-resolution Remote Sensing Image Instance Segmentation Based On Deep Learning

Remote sensing image instance segmentation is of great significance in extracting ground object information from aerospace photographic images.Compared with the traditional manual interpretation method,using computer vision technology to automatically detect and segment objects can greatly reduce the labor cost,and improve the efficiency.However,due to the high density,arbitrary shape and huge scale variation of objects in high resolution remote sensing images,the task of instance segmentation based on high-resolution remote sensing images faces great challenges.Therefore,it is very meaningful to study the instance segmentation methods of high resolution remote sensing image.This thesis first summarizes the existing research results in the field of instance segmentation,analyzes the shortcomings of the existing deep learning model for instance segmentation and the challenges faced by remote sensing image instance segmentation,and proposes two instance segmentation method for the problems faced by different applications.The main work of this thesis is as follows:1.Aiming at the problem that the existing instance segmentation methods can not effectively deal with objects with huge scale variation,an adaptive fusion strategy is proposed.This strategy adds a weighted full connection fusion mechanism between the "top-down" information flow path and "bottom-up" ones,which makes the feature information flow fusion more flexible and efficient,improving the segmentation performance.Aiming at the problem of insufficient segmentation accuracy of arbitrary shape objects,a content-aware upsampling method is proposed.This strategy can help the model accurately capture the differences between feature content and reduce the loss of feature information in the process of up sampling.Compared with the baseline method PANet,the performance of object detection and instance segmentation of our method is improved by 5.5% and 4.1% respectively.Among them,the performance of instance segmentation with arbitrary shape objects is improved significantly.The experiment suggest that the proposed method has obvious advantages in dealing with the problem of segmenting arbitrary shape objects and learning the features of objects with huge scale variation difficulty compared with the existing instance segmentation method.2.Aiming at the problem that dense objects are difficult to be detected and segmented effectively,an adaptive sample selection strategy is proposed,which constrains the proposal box to make the overlap ratio between proposal box and corresponding ground truth box higher.Moreover,the method adjusts the threshold used for determining positive and negative samples according to the difference of each real instance,which improves the quality of positive samples,and reduces the interference between adjacent instances in dense scenes.And it also solves the problem of detecting and segmenting difficulty for dense small objects.Compared with the baseline method PANet,the performance of object detection and instance segmentation is improved by0.4%.Especially,the performance of object detection and instance segmentation in dense scenes is improved.Experiment suggests that,compared with the existing instance segmentation methods,the proposed method has clear superiority in dealing with the problem of detecting and segmenting difficulty in dense scenes.