Remote Sensing Image Detection Based On Convolutional Neural Network

The task of remote sensing image detection is to categorize and locate objects in some given remote sensing image.Remote sensing image detection is widely used in military and civilian fields,and it is well worth the time and effort to research.After years of research and exploration by scientific researchers,natural image detection has become more mature.However,remote sensing object detection is difficult due to the particularity and complexity of remote sensing images.Therefore,for the remote sensing detection,the corresponding detection algorithm should be designed according to its characteristics.This paper mainly studies the remote sensing image detection based on convolutional neural network.First,this paper designs the FPN-RPN remote sensing image detection model based on horizontal bounding box regression.Based on Faster-R-CNN,this model combines ResNet50 with FPN as a feature extraction network,and selects RoI Align instead of RoI pooling.The RoI Align layer fixes the feature map to a uniform size,avoids the loss of accuracy in the quantization operation,and reduces the negative impact on the detection of small objects.During the model training,OHEM are used to balance the number of positive and negative samples to improve the robustness of the model.During the model testing,SoftNMS is used instead of NMS to post-process the region proposals and prediction bounding box which reduces the probability that the predicted bounding boxes of adjacent objects are too close to be suppressed,and improves the recall rate of detection.Aiming at the characteristics of the arbitrary direction of remote sensing images,FPN-RRPN remote sensing image detection model based on rotating bounding box regression is proposed.The model also uses ResNet50 combined with FPN to construct a backbone network for feature extraction.The extracted features are sent to the RRPN network to generate rotating region proposals to reduce the difficulty of rotating box regression and improve the detection accuracy of the rotating bounding box regression model.In the model,RRoI Pooling is used to resize the rotated feature map to a uniform dimension,which is necessary to input the subsequent fully connected layer for prediction objects classification and position refinement.In order to deal with the regression of the rotating bounding box,an IOU and a Non-maximum suppression algorithm suitable for the rotating box are adopted to further improve the detection accuracy of the rotating box regression model.In order to verify the effectiveness of the model proposed in this paper,three data sets of DOTA,HRSC2016 and UCAS＿AOD are used for testing.On the DOTA data set,the mean average accuracy of the FPN-RPN and FPN-RRPN models reached 65.3% and 65.7%,which exceeded the benchmark by 4.8% and 11.6%,respectively.In the UCAS＿AOD data set,the mean average precision of the FPN-RPN and FPN-RRPN models is as high as 89.6% and89.8%.In addition,the DOTA data set is used in the FPN-RPN model algorithm to test SoftNMS and multi-scale training and testing strategies.Experiments have confirmed that multiscale training and testing strategies can effectively improve the accuracy of target detection.However,the detection rate is significantly reduced at the same time.Since the DOTA data set contains more densely close objects,the use of Soft-NMS post-processing method can improve the recall rate of detection,and the mean average accuracy is improved by 1.1%.In the detection tasks of the DOTA and HRSC2016 data sets,the two models proposed in this paper have improved mean average accuracy compared to the benchmark algorithm.At the same time,they have also obtained excellent detection results on the UCAS＿AOD data set,which proves the effectiveness of the algorithm in this paper.