Single-Stage Remote Sensing Object Detection Based On Convolutional Neural Network

With the rise of deep learning research in image processing,the field of natural optical image processing has gradually transitioned from traditional methods to deep learning-based image processing methods,achieving significant performance improvement.This has attracted widespread attention in the field of remote sensing image processing.However,due to the fact that remote sensing images are acquired from a top-down perspective,allowing objects to appear in the image at any position and orientation,in contrast to natural optical images where objects are typically located at the center of the image with a fixed orientation,there is a substantial difference.Furthermore,remote sensing images can take various forms,such as optical images,Synthetic Aperture Radar(SAR)images,hyperspectral images,etc.,and the diverse nature of these images differs greatly from natural optical images.Consequently,it is not feasible to directly apply deep learning image processing methods developed for natural optical images to remote sensing images.Instead,research in deep learning image processing for remote sensing images needs to take into account the specific characteristics of remote sensing images.Remote sensing object detection,as one of the fundamental tasks in remote sensing image processing,is confronted with several challenges:(1)Multiscale issue: In natural images,objects are typically located near the center of the image with little variation in size.However,the extremely high capture altitude of remote sensing images results in significant scale differences even among objects of the same class,posing challenges for deep learning-based object detection.(2)object orientation variation: Due to the remote sensing images are imaged from a bird’s-eye view,objects can be oriented in any direction.Varied orientations within the same category complicate deep learning-based object detection.(3)Complex background: The wide imaging range of remote sensing images leads to diverse content and backgrounds,presenting a challenge for deep learning methods to effectively distinguish background information.(4)Dense distribution of objects: Objects in remote sensing images are often densely distributed in the same direction,especially in artificially planned areas such as vehicles in parking lots or ships in ports.The dense distribution of objects results in small inter-object distances,making deep learning methods prone to both missed detections and false detections.In addressing the challenges of remote sensing object detection,this dissertation conducts in-depth research on deep learning-based remote sensing object detection,focusing on improving the performance through features extraction,embedding prior information,and regression representation of remote sensing objects.The research content and innovative aspects of this dissertation are outlined as follows:1.In addressing the multiscale issue of objects in remote sensing images,this study initially investigates the feature extraction capability of remote sensing object detection networks.Specifically,a module with the capability is proposed to extract multiscale object information,which not only addresses the multiscale issue of object but also accurately locates their positions.Furthermore,the study explores the degree of information fusion between feature maps in the feature pyramid structure.A multiscale feature pyramid with adaptive feature fusion capability is designed to effectively capture information from different scales of objects.Through the investigation of the multiscale information extraction capability of remote sensing object detection networks,this research contributes to the enhancement of the performance of remote sensing target detection.2.In response to the challenges posed by the complex background,orientation variations,and dense distribution of remote sensing objects,this dissertation addresses the extraction of rotation-invariant features in the context of deep learning-based object detection methods.Specifically,the study incorporates the Local Binary Pattern(LBP)method into the feature extraction process of deep learning-based object detection.To address the issue of object orientation variations,the rotation-invariant Local Binary Pattern is implemented,endowing the deep learning-based object detection network with the capability to extract rotationinvariant features.By integrating the prior information of rotation-invariant Local Binary Patterns into deep learning-based object detection methods,a significant improvement in accuracy is achieved in remote sensing object detection.3.In addressing the challenges posed by changes in orientation,complex backgrounds,and dense distribution of objects in remote sensing images,this dissertation starts from the inherent characteristics of remote sensing image objects.A deep learning-based object detection method with direction prediction is employed to detect objects in remote sensing images.To overcome issues associated with angle-based arbitrary-oriented object detection methods,the dissertation proposes a novel ellipse parameters-based representation method for arbitrary-oriented objects,which conceals the angle within the focal vector of the ellipse,mitigating the problems associated with direct angle prediction.Furthermore,based on the proposed ellipse parameter-based representation method,the dissertation introduces a coarse-to-fine positive and negative sample selection strategy within the two-dimensional Gaussian distribution region of the object.Finally,experiments are conducted to demonstrate the effectiveness of ellipse parameter-based representation method.This approach provides a new perspective for research on arbitrary-oriented object detection.4.In response to the issues of angle periodicity and discontinuity at boundaries caused by direct angle prediction in arbitrary-oriented object detection in remote sensing images,this dissertation conducts an analysis of the advantages and disadvantages of the current vector decomposition-based representation methods for arbitrary-oriented objects.Based on this analysis,a more rational,simple,and efficient representation method for arbitrary-oriented objects is proposed,leading to a significant improvement in the performance of remote sensing image object detection.Additionally,an efficient algorithm for the conversion between angle representation and vector decomposition representation is introduced,reducing data processing time during training and accelerating the training of object detection networks.Furthermore,to achieve better detection on multiscale complex datasets,an adaptive positive and negative sample selection algorithm is implemented to alleviate the low positive sample matching rate for small objects.Through extensive experiments on multiple complex remote sensing image datasets,the effectiveness of the proposed vector decomposition representation and each module is validated in this dissertation.