Research On Object Detection Method In Remote Sensing Images Based On Deep Learning

With the rapid development of remote sensing technology,the quantity and quality of remote sensing images have been greatly improved.As a key component of remote sensing image analysis and processing,the development of remote sensing image object detection technology plays a vital role in promoting the development of civil and military science and technology.However,it is very difficult to detect objects in remote sensing images because of the complex background,large difference in the scale and shape of objects,and the influence of vertical vision.How to quickly and accurately detect typical objects in remote sensing images with complex environment and improve the accuracy and efficiency of automatic detection of remote sensing images has become an urgent problem in the field of remote sensing image analysis.Therefore,the research on remote sensing image object detection technology has high academic value and application value.The object detection method based on traditional technology in remote sensing images usually require manual extraction of object features,which is not comprehensive and accurate,and is time-consuming and labor-intensive.In recent years,due to the rapid development of deep learning technology,remote sensing image object detection technology has significantly improved the accuracy and efficiency of remote sensing image object detection by virtue of the powerful automatic feature learning ability of deep learning algorithm.However,due to the complexity of remote sensing images,there are still some key and difficult problems to be solved in remote sensing image object detection based on deep learning.Based on in-depth analysis of remote sensing images and the characteristics of objects in them,this paper proposes three exploratory object detection methods in remote sensing images combined with deep learning technology,which are detailed as follows:?Aiming at the problem that the scale of objects in remote sensing images is very different,and there are a large number of small objects,whose semantic information may be weakened or even disappear in the deep neural network,this paper presents an extended feature pyramid network(EFPN)and designs an adaptive strategy for object detection in remote sensing images.Among them,the EFPN is used to enhance the effective feature extraction ability of the network model.In the EFPN,a multi-branched dilated bottleneck module(MBDB)is designed to capture more multi-scale features and semantic information through the receptive fields of different branches.Then,an attention pathway is designed to better locate the objects.Finally,an augmented bottom-up pathway is designed to make the shallow layer information easier to spread and further improve the detection performance of small objects.In addition,an adaptive scale training strategy is designed to enable the network model to better recognize multi-scale objects in remote sensing images.At the same time,a clustering method is designed to realize the adaptive anchors setting,so that the network model can better learn the data characteristics of remote sensing images.?In view of the complex background of remote sensing images,most of which are from the top view,and there are many dense objects,the existing object detection network model based on deep learning still has the problem of missed detection and false detection.This paper proposes a region-attentioned network(RANet)with location scoring dynamic-threshold NMS to improve the performance of remote sensing image object detection under complex backgrounds and dense objects.Firstly,in order to detect objects in remote sensing images under complex background,a saliency constraint is introduced into the RANet,and a feature extraction network with regional attention is designed to effectively enhance the features of the object region.Then,considering that there are many dense objects in remote sensing images,a dynamic-threshold NMS method for overlap detection elimination is designed.Moreover,for solving the unmatching issue between the location accuracy and the classification score,the intersection of union detection header is further employed in the RANet to obtain the location confidence of the prediction box,and the final score of the prediction box was obtained by combining the classification score and the location confidence.Therefore,the location scoring dynamic-threshold NMS is designed,which can further improve the detection performance.?In view of the huge differences in scales and shapes of objects in remote sensing images,existing deep learning-based object detection networks are not ideal for this situation,and obtaining high-precision remote sensing image object detection performance usually requires a large and complex network structure that often contains a large amount of redundant information,which makes it difficult to achieve high detection efficiency,and it is also more difficult to deploy network models on resource-constrained devices for real-time application.This paper proposes an efficient remote sensing image object detection method based on PG-YOLO and optimizing channel pruning.In the PG-YOLO,a polymorphic module is first designed to simultaneously learn multi-scale and multi-shape object features in remote sensing images,so as to better detect objects with huge differences in remote sensing images;then,a group attention module is designed to make better use of the diversiform concatenation features in the network;finally,the multiple detection headers with adaptive anchors are designed to further improve the multi-category and multi-scale object detection performance of the one-stage detector PG-YOLO.In addition,based on PG-YOLO,this paper further designs an effective optimizing channel pruning method for model compression,which can significantly improve the efficiency of object detection to achieve real-time detection.In order to verify the effectiveness and advancement of the proposed method,the comparative experiments are conducted on several open source remote sensing image object detection datasets with the existing typical detection methods.The experimental results show that the proposed method is superior to the existing typical detection methods in terms of subjective effect,objective index and detection efficiency.