Research On Deep Convolutional Neural Network Based Object Detection Methods In Remote Sensing Images

Automatic,accurate and efficient detection of objects in remote sensing images is a fundamental but challenging problem faced for remote sensing image analysis.Recently,deep convolutional neural networks(DCNN)based algorithms has shown their much stronger detection power in computer vision field,which have achieved significant performance improvements over traditional methods.However,several challenges limit the applications of DCNN in objects detection from remote sensing images:(1)DCNN based object detection methods are particularly designed for detecting the bounding box of the targets without extracting attributes.(2)DCNN based methods with a fixed receptive field cannot match the scale variability of different objects and have poor localization performance with small objects.(3)The available annotated samples are not sufficient in number for training DCNN based methods.Therefore,most deep detectors have to fine-tune networks pre-trained on ImageNet that has little flexibility to redesign the network structure.(4)The DCNN model does not have the ability for modeling the rotation change.To address these problems,this dissertation systematically studies the DCNN based object detection methods in remote sensing image.Experiments results on differect object detection tasks from multi-source remote sensing images verify the effectiveness and practicability of the proposed method.The main innovations achieved in this paper are as follows:1.In the research on single-class object detection and attribute prediction,a coupled convolutional neural network based method is proposed for object detection and attribute prediction.In this method,the multi-task learning mechanism is used to collaboratively learn the loss function of attribute prediction with the object classification and bounding box regression.In addition,the multi-level feature fusion mechanism is adopted to improve the detection performance of small-sized targets.Our method combines the object detection network and attribute learning network in one network to extract the objects' location and attributes simultaneously,which expands the function of DCNN-based object detection methods.Comprehensive evaluations on the public Munich vehicle dataset and the collected vehicle dataset demonstrate the accuracy and effectiveness of the proposed method.2.In the research on multi-class object detection,this dissertation proposes a unified and effective method for simultaneously detecting multi-class objects in remote sensing images with large scales variability.Firstly,we redesign the feature extractor by adopting Concatenated ReLU and Inception module,which can increases the variety of receptive field size.Then,the detection is performed by two sub-networks: a multi-scale object proposal network for object-like region generation from several intermediate layers,whose receptive fields match different object scales,and an accurate object detection network for object detection based on fused feature maps,which combines several feature maps that enables small and densely packed objects to produce stronger response.The quantitative comparison results on the challenging NWPU VHR-10 data set,aircraft data set,Aerial-Vehicle data set and SAR-Ship data set show that our method is more accurate than existing algorithms and is effective for multi-modal remote sensing images3.Aiming at the limitation of network structure design of pre-trained ImageNet models and the large difference of domain distribution between remote sensing image domain and ImageNet image domain,this dissertation proposes an effective approach to learn deep object detector from scratch.Firstly,we design a condensed backbone network that consists of several dense blocks,which makes it easy to train.In addition,feature reuse strategy is adopted to make it highly parameter efficient.Therefore,the backbone network could be freely designed and effectively trained from scratch without using a large amount of annotated samples.Secondly,we improve the cross-entropy loss to address the foreground-background imbalance and predict multi-scale object proposals from several intermediate layers to improve the recall rate.Then,position-sensitive score maps are adopted to encode position information into each object proposal for discrimination.The comparison results on the Sentinel-1 dataset show that learning object detector from scratch achieved better performance than ImageNet pre-trained model based detectors.Our method is more effective than existing algorithms for detecting the small and densely clustered objects.4.In the research on arbitrary-oriented object detection and target recognition,a active rotation filter based object detetcion method is proposed for modeling the rotation change of the objects in remote sensing image.The method replaces the traditional convolution operation with the active rotation filters to capture the response characteristics of the objects in different directions for bounding box regression,and uses an oriented response pooling operation to extract rotation invariant features for target recognition.Based on the bounding box detection and recognition results,the main direction of each object is extracted by Hough transform line detection.Finaly,the arbitrary-oriented detection result is determined according to the identification model of the target.The quantitative comparison results on the collected challenging NUDT-Ship data set show that the object detection method using active rotation filter has higher recognition accuracy,and the arbitrary-oriented detection result can effectively distinguish dense objects.