Research On Object Detection Methods Of Optical Remote Sensing Image Based On Convolutional Neural Network

In recent years,due to the rise of deep convolutional neural networks,image target detection has been greatly developed,bringing huge breakthroughs to many related applications.In the field of computer vision,the object detection of optical remote sensing images has always been one of the important and extremely challenging tasks.Compared with natural scenes images,objects in remote sensing images are mostly characterized by diverse scales,dense arrangement,uncertain direction and complex background,which cause great difficulties in classification and positioning tasks.This paper studies the object detection method of optical remote sensing image based on deep convolutional neural network from two aspects of feature extraction and object representation form,which solves the problem of insufficient performance of existing target detection technology in the field of remote sensing.Image feature extraction is a very important step of object detection,which is directly related to the detection effect.Particularity for remote sensing image to extract object feature is difficult to make a clear distinction,this paper proposes a remote sensing image object detection method based on multi-FPN.The backbone network of this method uses residual network,connected by lateral connection and cross layer composition characteristics of the pyramid structure,will be to extract the feature information of multi-scale fusion,the spread of enhanced features and reuse.In addition,by using Soft-NMS instead of the original NMS method,the post-processing method is improved to reduce the false negative rate.The experimental results show that the improved method in this paper has good detection performance and the average detection accuracy reaches 89.42% on the remote sensing public dataset set labeled with horizontal frame NWPU VHR-10.The result of object detection is not only affected by the characteristic expression,but also related to the form of object representation.Most optical remote sensing image object with some rotation angle,using rotating bounding boxes that more joint object shape are relatively accurate,but there is limited by angle representation in nearly horizontal object detection,lead to regression loss a sharp rise,It is difficult to learn the correct objects location,precision is worse than horizontal bounding box methods.Aiming at the respective problems of the two target representations,this paper proposes a remote sensing image object detection method based on discriminator classification,which simultaneously predicts the horizontal and inclined bounding boxes in the detection stage.The area ratio discriminator was used to detect near horizontal objects with horizontal bounding box,and objects with some rotation angles were detected with inclined bounding box,and INMS was used to reduce the repetition boxes of the same object.Finally,by combining multi-FPN and discriminator classification detection network,extensive experiments on the remote sensing public dataset labeled with rotating frame DOTA.The mean average precision reaches 62.40%,with excellent performance.