Optical Remote Sensing Image Object Detection Based On Object Candidate And Multi-Feature Fusion

In the research of remote sensing image interpretation,object detection has always occupied an important position.The continuous development of remote sensing technology has prompted more and more high-resolution optical remote sensing images to appear in people’s eyes.Object detection in optical remote sensing images is widely used in both military and civilian fields.In recent years,deep learning has developed rapidly and is used in various computer vision tasks.Due to its powerful feature expression capabilities,convolutional neural networks perform very well in both natural scene image classification and object detection tasks.At present,some progress has been made in the application of convolutional neural networks to object detection task in optical remote sensing images.This thesis proposes three object detection methods to address some problems faced in current optical remote sensing image object detection,such as wide variation range of object scales,complex image background,low contrast and diverse and complex object appearance.Aiming at the difficulty of detection caused by a wide variation range of object size in optical remote sensing images,this thesis proposes a remote sensing image object detection method based on multi-scale feature fusion.It is a two-stage object detection algorithm,which is based on the VGG16 model and uses a Region Proposal Network(RPN)to generate candidate regions.Through using a top-to-bottom connection to fuse features at different levels,this method increases the adaptability of the network to objects with different scales,therefore improving the accuracy of object detection.In addition,in the head detection network,candidate region features on feature maps at different levels are also fused.This fusion method combines the details and neighborhood information of the object to provide clues for object detection.In order to avoid the problem of over-fitting caused by insufficient training data during training,the algorithm horizontally flips images to enhance the data set,simultaneously using transfer learning method for network training.Focusing on low object detection accuracy caused by the complex background and low contrast of optical remote sensing images,this thesis proposes a remote sensing image object detection method based on attention modules.It is an improvement on the multi-scale feature fusion object detection method in(1).There are two points for the improvement.First,a multi-branch convolutional layer is designed.The multi-branch convolutional layer uses convolution kernels of different sizes to process input feature maps in parallel and combines the processed results,which increases the width of the network and improves its ability to express features.Second,multiple attention modules are used,including spatial attention module,channel attention module,and pixel attention module.Attention modules weight optical remote sensing image features.This can reduce noise interference,improve the saliency of objects and guide the network to learn objects,thereby improving object detection accuracy.Aiming at the problem that diverse and complex object appearance leads to the difficulty of detection,and the low object detection accuracy caused by the insufficient understanding of geospatial object spatial structure information,this thesis proposes a remote sensing image object detection method based on multi-model decision fusion.In the method,a contextual information fusion sub-network fuses both local contextual features and object-object relationship contextual features to effectively identify objects and reduce false detection between different types of objects with similar appearance.A multi-region feature fusion subnetwork fuses multiple partial features of a object,which can mine more details of the object and increase semantic information,therefore improving detection accuracy.In addition,this method also designs a multi-model decision fusion strategy to fuse outputs of multiple subnetworks for obtaining final consequences.The strategy can make the algorithm more robust and achieve better detection results.