Search Of Remote Sensing Image Target Detection Based On Depth Neural Network

In recent years,remote sensing technology has developed rapidly.How to efficiently and accurately use remote sensing data for target detection is the current research focus of remote sensing technology.Remote sensing image target detection refers to the identification and positioning of the target object in the remote sensing image,which can provide the basis for subsequent remote sensing information processing and application,and is an important part of current remote sensing processing.In remote sensing images,the background is complex,the environment is changeable,and the performance of traditional target detection methods is not good.In recent years,the rapid development of deep neural networks has provided new ideas for remote sensing image target detection.The target detection algorithm based on deep neural network can extract the deep essential features of the image through a large number of supervised training samples,and has strong generalization ability and excellent detection effect.For the field of remote sensing images,the available training samples are scarce.Among the current mainstream target detection algorithms based on deep neural networks,one type relies on a large number of target-level training samples and is difficult to be directly applied on remote sensing images;the other type relies on a large number of image-level training samples,with insufficient prior information in candidate regions and inability to detect frames Adaptive problem.This article takes the detection of aircraft targets in remote sensing images as an example to study how to use fewer samples to better complete target detection in remote sensing images.It mainly includes the following parts:(1)Aiming at the problem of lack of samples in the detection of aircraft targets in remote sensing images,a detection model based on deep neural network DC-DNN(Density clustering with deep neural network,DC-DNN)is proposed,which can be completed by only a few image-level samples Aircraft target detection and segmentation tasks.First,use the bottom features of the image to make pixel-level labels to complete the full convolutional neural network(FCN)model training,and combine the FCN model with the DBSCAN density clustering algorithm to select the adaptive candidate area of the aircraft target;then,extract based on the VGG-16 network The high-level features of the candidate area are classified to obtain the aircraft target detection frame;finally,the detection framesuppression algorithm is used to eliminate the overlap frame and the false detection frame at the same time,and the final aircraft target detection result is obtained.The experimental results show that the DC-DNN model has reached 95.78%,98.98%,and 0.9735 on the three evaluation indexes of accuracy,recall and F1 score,respectively,and has excellent detection performance.(2)Aiming at the problem of repeated feature extraction and low detection efficiency in the DC-DNN model,the Ro IPooling layer is introduced into the RDC-DNN(Ro IPooling fusion with DC-DNN,RDC-DNN)model proposed in the DC-DNN model.First,introduce the Ro IPooling layer to merge the feature extraction network of FCN and CNN in the DC-DNN model to unify feature vectors,reduce classification difficulty,and improve detection efficiency;then,remove the candidate region offset expansion and detection frame suppression in the DC-DNN model Operation to improve detection efficiency.Experimental results show that the detection speed of the RDC-DNN model is increased by25.98% compared with the original model,which greatly improves the detection efficiency and verifies the effectiveness of the model.