| In recent years,with the rapid development of UAV technology,satellite plat-form,sensor technology and other technologies,people can obtain a variety of types of remote sensing image data with richer information.Scientific,rational,and efficient processing and utilization of the rich information contained in remote sensing images can create tremendous application value for fields such as agricultural monitoring and evaluation,military reconnaissance,high-precision terrain mapping,geographic infor-mation systems,traffic control and other fields.In remote sensing image processing,object detection is one of the most fundamental and important tasks.The research of improvement high-performance upon the object detection methods in remote sensing images have significant theoretical and great application value.There are two main tasks in remote sensing image target detection:one is to locate the position of var-ious objects on the remote sensing image;the other is to locate the target category recognition.Traditional methods of detecting targets in remote sensing imagery often rely much heavily on people’s prior experience,which results in limitations for the insufficient gen-eralization ability of the detection methods.In recent years,with the development of Convolutional Neural Network（CNN）,Deep Learning（DL）and other technologies,the research on object detection methods based on CNNs has made significant progress and has increasingly gained attention from people.At present,research on object detection methods based on neural network and deep learning has become one of the mainstream methods for remote sensing image object detection.In general,the methods of object detection can be classified into two basic methods:one-stage and two-stage frameworks.However,in practical applications of remote sensing image object detection,several chal-lenging problems still exist,including:（1）Remote sensing images often have complex backgrounds and a large coverage scope,which makes it difficult to processed them as a whole,and need to process on the pixel-level block;（2）Remote sensing images often adopt Bird′s eye view,so the targets are often small,densely arranged,and the target orientation is arbitrary.More,there is a significant difference in size between different categories of targets.（3）The receptive field of remote sensing images often has a signif-icant impact on convolutional neural networks;（4）remote sensing images often contain different types of noise,leading to insufficient robustness in target detection methods and limiting their generalization ability.This paper focuses on the aforementioned challenging problems and studies research on remote sensing image object detection methods more.Based on convolutional neural networks and deep learning technologies,several new object detection methods are proposed.Experimental comparative analysis is conducted on multiple public datasets for remote sensing image target detection,which verify the effectiveness of the proposed methods.The main research content and innovative achievements of this paper are as follows:（1）A two-stage rotation object detection method based on the scale factor loss function is proposed.To address the challenges of complex background,different types and scales of objects,small and densely arranged objects and arbitrary orientation in aerial remote sensing imagery,firstly,the concept of scale factor was proposed through analyzing the detection errors of objects according to the theory of statistical regression error.Subsequently,a new regression loss function,namely Smooth L1loss function with scale factor,is proposed.Finally,based on Smooth L1loss function with scale factor,using Faster R-CNN as the baseline network and Res Ne Xt as the backbone network,a new two-stage rotation object detection method is proposed.In addition,to solve the problem of arbitrary target orientation,we have embedded an Ro I-Transformer module,which can help network learn the rotated target candidate region from the regular rectangular candidate region.Relevant experimental analyses were carried out on two well-known remote sensing object detection datasets,DOTA and HRSC2016.The experimental results show that this method can effectively and stably improve the detection accuracy of small targets.Compared with other major object detection methods of the same type,this method has better performance in object detection.（2）A single-stage rotation object detection method based on a dual-layer semantic information balancing module is proposed.Most networks use feature pyramid structures to fuse feature information from top to bottom in remote sensing images to alleviate the impact of different scales of target types and the same type of targets.However,this feature fusion method often leads to the fusion of too much shallow information from high-level semantic information.In order to balance the characteristic semantic information from different scales layers,this paper proposes a new plug-and-play module,which is a dual-layer semantic information balance module（DBFP module）based on the semantic information balance module（BFP module）.The DBFP module enhances the semantic information of shallow layers,it not only makes high and shallow-level semantic features distinctive,but also improves the detection performance of small objects.Furthermore,we continue to introduce a scaled smooth L1loss function with a scale factor to achieve overall scale normalization of the targets.Experiments show that the proposed method is effective and stable,and the overall performance of object detection has been improved as well.（3）A rotation object detection method based on self-adjusted scale factors has been proposed.Based on the previous research,and taking into account the characteristics,we use sample images in a batch during the training phase,and propose a concept of self-adjusting scale factors based on statistical regression error theory.Subsequently,a new regression loss function,self-adjusted Smooth L1loss function with self-adjusting scaling factor,is proposed.Afterwards,the existing one-stage and two-stage primary object detection methods were improved based on this loss function.In addition,the dual-layer semantic information balancing module（DBFP module）continued to be applied.In addition,the dual-layer semantic information balancing module（DBFP module）was also continued to be applied.The experimental results indicate that the proposed method is effective and stable.At the same time,the overall performance of object detection has also been improved. |
