Object Detection And Instance Segmentation In Optical High-Resolution Remote Sensing Imagery Based On Mask R-CNN

Automatically or semi-automatically detecting objects from high-resolution images is an important research topic in remote sensing field.Optical aerial and satellite images with higher spatial resolution have smaller feature gaps with the natural images.Therefore,it is possible to apply the DL algorithms to remote sensing image recognition.Mask R-CNN is the most state of the art DL model which has a powerful capability of object detection and instance segmentation in parallel.It has achieved promising progress on natural image recognition.Within the framework of this master thesis,we apply the state of the art technologies and techniques in computer vision and DL domain to remote sensing area.The objective is to investigate the generalization of model Mask R-CNN in remote sensing field.Moreover,since remote sensing data is an important source of vector maps,the model is expected to find missing instances on the map and improve the quality of vector maps.In practice,there is a large amount of remote sensing images and needs for real-time analysis.Therefore,improving the training speed is also of great importance.We first conduct a multi-class experiment which detects and simultaneously segments the sports fields(baseball diamond,basketball field,ground track field,stadium,tennis court).Besides,by replacing the deep backbone network to a shallower one,we intend to weight the tradeoff between overall accuracy and training speed for a relatively small remote sensing dataset.Moreover,we trained another single-category recognition Mask R-CNN model on building dataset.Additionally,it is tested with 3 different building density(dense,medium and sparse)datasets.In sports field task,the model achieves excellent object detection performance with 0.956(ResNet101)/0.957(ResNet50)mAP and instance segmentation results with 0.887(ResNet101)/0.865(ResNet50)mIU.For relatively small remote sensing dataset,shallow network can also attain comparative performance as deep network but saves much less time.The model shows a quite good generality from natural image community to optical high-resolution remote sensing domain.Moreover,there is only a small gap between object detection and more difficult instance segmentation results.Further visualization on extra test images and Shenzhen high-solution images both shows excellent results.It shows robustness to different pavements,various scales,varied shapes,multiple directions and image distortion.In building detection and segmentation mission,the model achieves quite good object detection performance with 0.797(ResNet101)mAP and instance segmentation results with 0.695(ResNet101)mIU.Additionally,the results tested on residential with different densities show a relatively good performance on sparse and medium residential but not quite good on dense residential.