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

Object detection is to locate and classify objects from images and is an important branch of image processing.With the rapid development of various ground observation detection equipments such as remote sensing satellites,the data of remote sensing image has drastically increase.Remote sensing image target detection can effectively interpret remote sensing images,and it is applied in many aspects,such as military deployment,target tracking,smart city construction,agricultural disaster detection and early warning,and car unmanned driving.Because of the remote sensing image has a large format and a wide coverage.The target size is different,the small targets are densely arranged and the color and shape of the target may be similar to the background,so it is difficult to detect the target on the remote sensing image.At present,there are many target detection models based on deep neural network,but these models are mainly applied to ordinary optical images.However,the training of such models requires a large number of training samples with labels,and it is difficult to provide such a large number of remote sensing image samples with labels in practice.Therefore,the existing target detection model is directly used in the remote sensing image for detection,and the detection performance and detection results are not ideal.Considering the above research status,based on the theoretical basis of the convolutional neural network,the thesis focusing on these difficults which existing object detection models in the remote sensing image data sets:First,detection performance of he small objects in the remote sensing image is poor,how to improve the detection capability of multi-scale object detection in complex backgrounds,especially for small targets;second,aiming at the problem that the marked remote sensing image samples are few,the model has poor target detection performance for remote sensing images with different feature distributions.Combining the target detection model and domain adaptation,the target detection model has higher detection performance in remote sensing images with large differences in the distribution of image characteristics such as illumination and visibility,so that the model has better Promotion ability.The main research of this paper is follow as:Aiming at the problem that the detection performance of the model decreases significantly when the training data and test data are significantly different and the limited training samples,this paper proposes Fusion features based deep learning remote sensing image target detection model.The model uses a small-scale network structure and proposes a strategy of fusing multi-level features to obtain more effective features,which enables the model to improve the detection accuracy of denser and different-scale targets in remote sensing images without increasing the detection time.At the same time,a new post-processing algorithm—Packet Fusion Reject detection bounding boxes is proposed to remove redundant frames,remove redundant detection frames,and fine-tune the position of the detection frame to make the detection frame locate the target more accurately and further improve the detection accuracy.The experimental results show that the model reduces the missed detection rate and false detection rate while improving the detection accuracy,and achieves high detection performance on small-scale and dense targets.Aiming at the problem that the convolutional neural network needs a large number of labeled samples to train the model,when the difference between the training data and the test data is large,the detection performance of the model decreases significantly.This paper proposes Skip-Feature Pyramid Domain adapted model for remote sensing image target detection.In order to improve the problem that the object detection of the model is not fit to the task,the domain adaptive structure is introduced into the detection model.Through the training of the domain structure,the feature extraction ability of the model on different remote sensing images is improved,and the robustness of the model is improved.Secondly,the feature map generated by the convolution is up-sampled step by step using bilinear interpolation,and the feature map generated by the convolution of the lower layer is pixel-wise added and fused to obtain a multi-scale feature map.At the same time,a skip connection information stream is added to the upsampling structure.The jump connection information flow is a direct fusion of high-level feature maps and lower-level feature maps,so that the semantic information of low-level feature maps is more.When extracting candidate regions,the RPN is used to predict the candidate regions on multiple different feature maps,so that the candidate region extraction is more effective and the target detection performance is better.The model in this paper achieves the best detection accuracy in the NWPU-VNR10 and RSOD-DATA data sets compared with the detection results of other comparison detection models,and has the lowest number of missed and misdetected cases.