Instance Segmentation Of High-resolution Remote Sensing Images Based On Deep Learning

With the rapid development of remote sensing technology and imaging technology and the rapid expansion of high-resolution remote sensing image data capacity,automatic object detection and instance segmentation of high-resolution remote sensing images have become a research hotspot of current remote sensing applications.Traditional automatic object detection methods of remote sensing images are mainly based on artificially defined features.However,in the instance segmentation of high-resolution remote sensing images,artificially defined features are hard to achieve strong feature description capabilities,and new theories and methods are urgently needed.In recent years,deep learning theory has made important breakthroughs in automatic feature extraction and end-to-end feature learning,and object detection in remote sensing images based on deep learning has attracted extensive attention.However,compared with visual images,remote sensing images have the characteristics of complex background,diverse targets,and clutter interference.Typical deep-learning-based instance segmentation methods encounter difficulties of inefficiency or low precision when transferring from visual images to remote sensing images.In response to the above problems,this thesis combines deep learning theory with the characteristics of high-resolution remote sensing images,and focuses on three directions: the construction of remote sensing image dataset,the development of instance segmentation methods for optical remote sensing images,and the development of instance segmentation methods for SAR remote sensing images.The main work and innovations of this thesis are summarized as follows:1.Aiming at handling the shortcomings in existing SAR ship datasets,a high-resolution SAR image dataset（HRSID）was constructed through the steps of high-resolution raw SAR image acquisition and preprocessing,dataset annotation,dataset structure analysis,and dataset benchmarking.HRSID is widely applied in designing the ship detectors and instance segmentation networks in the SAR field.2.In order to construct a lightweight network for instance segmentation of optical remote sensing images,a lightweight adaptive Ro I extraction network（ARE-Net）is proposed.ARE-Net adopts global attention Ro I extractor and perceptual Ro I extractor to alleviate the impact of low sensitivity to physical deformation of optical remote sensing images on instance segmentation.The experimental results indicate that ARE-Net achieves the highest AP and the lightest network structure compared with the mainstream methods.3.For the sake of the low instance segmentation AP caused by the complex background and dense distributed small objects in optical remote sensing images,a consistent proposal regions of instance segmentation network（CPISNet）is proposed.CPISNet adopts the adaptive feature extraction network and ensures the consistency of mask predictions during training and testing process of instance segmentation through a cascade structure to improves the AP.Experiments on the i SAID and NWPU VHR-10 datasets indicate that compared with the mainstream methods,CPISNet achieves a significant improvement in AP,and suppresses the issues of false alarms,missed detections,and aliasing masks that appear in the mainstream methods.4.Aiming at solving the low AP caused by adjacent distributed ships and small ships in SAR images,a low-level feature-guided instance segmentation network（LFG-Net）is proposed.LFG-Net starts from the perspective of small target segmentation and combines super-resolution technology to build network modules.Experiments on the SAR datasets HRSID and PSeg-SSDD indicate that LFG-Net produces a huge improvement in instance segmentation AP compared to mainstream methods,and in complex scenes such as dense distributed small ships and adjacently distributed ships in the inshore area,LFG-Net can effectively segment the SAR ships.