Research On Remote Sensing Target Detection Technology Based On Convolutional Neural Network

With the development of China's space industry and remote sensing satellite technology brilliantly,the resolution of optical remote sensing images has been greatly improved,and image information has become rich.It helped the application of optical remote sensing images in agricultural production,environmental monitoring,military reconnaissance and other fields greatly.At the same time,the accompanying information is the diversity and complexity of information processing.Extracting key remote sensing image information becomes even more important.Object detection plays a role as an important method and one of the main research hotspots in remote sensing images.Especially in recent years,the popularity of deep learning has helped open up a wider path in the field of machine vision.Among them,the development of object detection algorithms based on convolutional neural networks is constantly changing.Compared with traditional object detection algorithms,it has stronger learning ability and generalization ability,and has played a more efficient role in the processing of natural images.Therefore,this article uses two different deep learning-based object detection methods for object detection in optical remote sensing images.The main research contents are as follows:(1)Aiming at the problem of poor detection of small targets in the single-stage end-to-end target detection algorithm represented by YOLO,we adopted the method of dense connection and multi-scale feature fusion,which mainly borrowed from the feature pyramid network.Multi-scale connection,and for the down-sampling process that easily loses feature information,the Inception structure is used to implement multi-channel dimensionality reduction,which further retains feature information.In order to make the network still meet the real-time performance,we have deleted some network layers and used dense connection methods for feature transfer.Without increasing the amount of parameters,the acquisition of deep features is guaranteed.Finally,we used the ship dataset in the Kaggle competition for training and testing,and clustered the sizes of the ships in the dataset to set the size of the generated anchor box,so that the network could converge faster.Compared with the original YOLO algorithm,the accuracy is improved,and the detection speed can reach 23 FPS.(2)A target detection algorithm based on a rotation detection frame is proposed.Its framework is a two-stage detection model based on classification and regression of target candidate regions.Compared with direct regression detection algorithms,it has higher detection accuracy and scalability.Sex.And we mainly do the rotation operation on the target detection frame,which includes the rotation region candidate network,the design of the rotation anchor point frame,the rotation operation of the region of interest pooling,etc.In addition,we conducted a comparison test on different kinds of pre-feature extraction networks,selected Inception-Res Net V2 with the highest accuracy as the backbone network,and added a multi-layer upsampling module to achieve the fusion of the underlying features and higher-level features,and The feature maps of the fusion of different layers achieve the parallel detection method.The redundant detection results are removed by non-maximum suppression,which further improves the detection accuracy and the IOU value of the detection frame and the real target.After that,we built a dataset of ten types of target scene rotation rectangle labeling datasets.On this dataset,we tested the improved network and compared the different rotation rectangle detection frame algorithms,which confirmed the rationality of the algorithm and reliability.