Research On Deep Learning-based Object Detection Algorithm For Remote Sensing Image

With the rapid development of remote sensing technology,the resolution and coverage area of remote sensing image data have been significantly improved.As a result,remote sensing image has become one of the most important means of obtaining information about the Earth’s surface.Remote sensing image object detection(RSOD)plays a core part in remote sensing image processing,is a fundamental technology for many applications.The main task of RSOD is to identify and locate the object of interest in remote sensing images,which is widely used in civil and military applications such as urban planning,traffic monitoring,land management,intelligent monitoring and military reconnaissance.In recent years,great breakthroughs have been made in object detection with the rapid development of deep learning.However,due to the special imaging characteristics of remote sensing images,RSOD still faces many challenges,such as the high resolution of remote sensing images resulting in high hardware consumption,noisy labels are prone to be introduced when labelling remote sensing image data for object detection tasks,the complexity of the background of remote sensing images and the high interclass similarity of remote sensing objects,which all make it difficult for RSOD algorithms to be applied in practical scenarios.Therefore,the paper improves the performance of remote sensing object detection algorithms and promotes the algorithms towards practical applications from the perspective of reducing the hardware consumption of high-resolution image,decreasing the negative impact of noisy labels,and designing two better object detectors,based on the characteristics of remote sensing images and analysing the challenges of applying the RSOD algorithms in practice.The main contributions and innovative works of this thesis are summarized as the following three parts:1.In order to address the problem of high hardware consumption caused by high-resolution of the remote sensing images,this paper proposes a superpixel-based processing algorithm for high-resolution remote sensing images.The common processing algorithms are to downsample a high-resolution image into a low-resolution image or to cut a high-resolution image into multiple low-resolution sub-images,while these processing methods still have some problems.The processing method of downsampling will lose the rich details of the object,which provided by the high-resolution remote sensing images,resulting in deteriorating the performance of the subsequent detection algorithms.The processing methods of cutting the high resolution into multiple sub-images are usually categorized into no overlapping region between sub-images and having overlapping region between sub-images.The cutting method of no overlapping between sub-images may lead to the object at the boundary of the sub-image being cut into multiple parts,which may destroy the integrity of the object,and then affect the performance of the detection algorithms.The cutting method with overlapping regions between sub-images can ensure the integrity of the object,which will generate too much redundant information.To address the above problems,this paper proposes a high-resolution remote sensing image processing algorithm based on superpixel.First,the high-resolution image is processed by the superpixel segmentation algorithm,and then it is cut into multiple non-overlapping low-resolution sub-images,and the superpixels are utilized to ensure the integrity of the object during the cutting process,which also keep the rich detail information of the high-resolution.Compared with the conventional way of processing high-resolution images,the proposed method not only preserves the integrity of the objects,but also reduces the redundancy information.The proposed method provides more efficient inputs to the subsequent deep learning-based RSOD algorithms,significantly improves the training and testing efficiency,and reduces the hardware consumption of a single input of the image in practical applications.2.In order to address the problem of the noise labels in remote sensing images affecting the performance of the RSOD,this paper proposes a robust training method for object detectors in remote sensing image.The quality of the labels plays a crucial role in the performance of the deep learning algorithm,and the accuracy of labels directly affects whether the deep learning algorithm can learn good weighting parameters.Due to the heavy and costly task of labelling remote sensing image data,the actual data labelling process is prone to introduce noisy labels,which affects the detection performance of the algorithm during training.Several approaches have been developed to learn robust models from noisy labels,but most of them have focused on classification tasks.Training object detectors with noisy remote sensing data has been less investigated.This paper proposes an object Co-teaching training strategy to train robust object detectors from noisy labels.Specifically,the object Co-teaching training strategy trains two detectors in a parallel manner,and lets them teach each other to filter noise object instances in each given mini-batch.Moreover,this paper also proposes a Reweighting object Co-teaching training strategy to improve the performance of the detector on clean datasets without adding any complexity of the detector.The proposed training strategy provides a guarantee to the algorithm when applied to real scenarios.3.In order to address the problems of complex background and high inter-class similarity of the remot sensing image,this paper proposes two novel RSOD algorithms based on YOLO.Due to the imaging characteristics of remote sensing images,it results in the complex background and high interclass similarity of objects,which makes a great challenge for the RSOD.In this paper,YOLO-DA and Graph-YOLO are designed for different application scenarios in response to the challenges encountered when detecting objects in remote sensing images.Firstly,YOLO-DA is designed for remote sensing application scenarios with limited computational resources.Specifically,an attention module at the end of the detector is designed for guiding the neural network to extract more efficient features from the complex background while also minimizing the amount of additional computation.Moreover,a lightweight decoupled detection head with enhanced classification and localization capability is developed to detect objects with high interclass similarity.The proposed YOLO-DA achieves a good trade-off between detection accuracy and detection speed,and has high practical application value.Then,the Graph-YOLO is designed for hardware-rich application scenarios.It designs an attention-based convolutional module to address the problem of complex backgrounds,introducing attention mechanisms with orientation-aware mechanisms at multiple locations of the network to improve the representation of features while avoiding interference from complex backgrounds.In order to solve the problem of high interclass similarity,Graph-YOLO adopts a decoupled detection head with better performance to identify similar object.In addition,Graph-YOLO combines graph neural networks and convolutional neural networks to extract the intrinsic topology of remote sensing images.This further improves the performance of remote sensing object detection algorithms,and achieves more accurate results of object detection.