Object Detection Of Remote Sensing Image Based On Improved Rotational Region Convolutional Neural Network

Remote sensing image target detection plays an important role in civil,military and other fields.However,due to its large size,small and dense targets at any angle,the targets easily blocked,the uneven target categories,and the complex background,target detection of remote sensing images is still a very challenging task.In recent years,the remote sensing image target detection method based on deep convolutional neural network has received more and more attention because of its advantages of high precision and fast processing speed.In this paper,the text detection method R2CNN（Rotational Region CNN）is introduced into the target detection field of remote sensing images.Combined with the characteristics of remote sensing images,the FM-R2CNN（Feature-Fused and Masked R2CNN）target detection model is proposed.The detection model considers several aspects such as feature extraction,introduction of mask information,and design loss function.It starts to solve the problems of objects with arbitrary angles,small and dense objects,occlusion of objects,and unbalanced sample categories in target detection of remote sensing images with a view to providing core technology for automatic analysis of remote sensing images.The main research contents and contributions of the paper are as follows:（1）In terms of model structure design,the low-resolution feature map and highresolution feature map extracted from the network are adaptively weighted and fused to reduce the loss of detailed information.At the same time,considering that most remote sensing targets are rigid bodies,it is easier to obtain segmentation labels.Therefore,the corresponding segmentation mask information is introduced as a training branch of the detection network according to the category of the remote sensing target to improve the accuracy of the target detection frame.（2）Remote sensing data comes from a wide range of sources,collected by different sensors.The data collected by different sensors is distributed differently.In order to make the model more robust,label smoothing is introduced to achieve regularization of the model,which improves the detection performance of the test set.（3）In order to solve the problem of imbalance between positive and negative samples for remote sensing target detection,in designing the model loss function,a new loss function is proposed for model training to improve the detection accuracy of small sample categories.（4）Through a large number of ablation experiments,the feasibility of various improvement strategies proposed in this paper is verified.And compared with the current advanced research methods,the mAP value evaluated by the method in this paper on the public data set DOTA reached 73.19%,which has certain comparative advantages.