Research On Methods For Partially Occluded Object Detection In Optical Remote Sensing Images Via Regions With CNN Features

Object detection has been a basic but challenging task for remote sensing images interpretation.With the great development of remote sensing technology,more and more attention has been paid tothe research of remote sensing images.There are two difficult issues that need be solved: how to quickly find as many suspicious objects at a small size as possible from complex backgrounds and how to detect partially occluded objects effectively in real remote sensing applications.Although the performance of general object detection has been improved dramatically by the CNN-based methods,the state-of-the-art object detection algorithm renders unsatisfactory performance as applied to detect small or partially occluded objects in remote sensing images.To address these issues,our efforts should be made in three aspects including data pre-processing,the network architecture of CNN-based feature extractor and loss function.The main contents and innovative points of this dissertation are as follow:1.Object detection based on Fast/Faster RCNN employing fully convolutional architectures and skip-layer connections is proposed.Based upon analyzing the feature extractor and the region-wise object classifier in Fast/Faster RCNN,we deduce the conditions of the input size in the fully convolutional networks containing the parallel branch of pooling layer and convolution layer.We argue that the input size has to be fixed in a cartain range when these fully convolutional networks are plugged into Fast/Faster RCNN.Furthermore,the idea of skip connection analogous to the hybrid of PVANET and FPN is presented to combine several intermediate outputs in the feature extraction stage.Consequently,the low-level visual features and high-level semantic features can be taken into account at the same time.Finally,we demonstrate that carefully designing the deep networks for region-wise object classifier can lead to good detection accuracy,which shows a profound understanding of the two-stage object detector.2.For small object detection in optical remote sensing images,a modified network of shared convolutional features is proposed.Firstly,the top-down modulation network is proposed as a way to incorporate fine details into the detection framework inspired by the human visual pathway.Secondly,the misalignment problem has a severe negative effect for the small object detection performance when RoIPool performs coarse spatial quantization for feature extraction.To fix this,we propose a simple,quantization-free architecture,called Context RoIAlign,that faithfully preserves exact spatial locations and context informations.Finally,a sampling strategy named Balanced-Sampling is presented to address the issue about the imbalanced numbers of images between different classes while a simple yet effective approach,namely Random Rotation,introduced to augment our available optical remote sensing data.Experimental results show that our modified network improves the mean average precision by a large margin on detecting small remote sensing objects.3.For partially occluded object detection in optical remote sensing images,a deformable Faster RCNN with aggregating multi-layer features is proposed.The region-based convolutional networks have shown their remarkable ability for object detection in optical remote sensing images.However,the standard CNNs are inherently limited to model geometric transformations due to the fixed geometric structures in its building modules.To address this,we integrate a new module named deformable convolution into the prevailing Faster RCNN.By adding 2D offsets to the regular sampling grid in the standard convolution,it learns the augmenting spatial sampling locations in the modules from target tasks without additional supervision.In our work,a deformable Faster RCNN is constructed by substituting the standard convolution layer with a deformable convolution layer in the last network stage.Besides,More TDM networks are applyed to not only generate more anchors to improve the recall of the region proposals but also generate more abundant feature maps with high semantic information at low layer,which can boost the small object detection performance greatly.To make the model robust to inter-class occlusion,a simple yet effective data augmentation technique called Random Covering is proposed for training the convolutional neural network.To keep each proposal away from the other ground-truth objects as well as the other proposals approaching different designated targets,the Smooth＿L1 loss function is substitued by Repulsion loss function in the bounding box regressor of RPN stage,which is experimentally verified to reject redundant proposals effectively.In the stage of region-wise feature classification,OHEM is aplplyed to automatic select hard examples,which can make training more effective and efficient.The experimental results show that the proposed method achieves obvious performance improvement on three remote sensing datasets...