Research On Aircraft Detection Technology Of Remote Sensing Image Based On Convolutional Neural Network

Since the Gulf War,remote sensing satellites have played an important role in modern local warfare as an important means of military reconnaissance.Target detection for remote sensing images has also become one of the hotspots in the field of military reconnaissance research.At present,the traditional target detection algorithm has been difficult to adapt to the processing requirements of high-resolution remote sensing images due to its low timeliness and weak generalization.The target detection algorithm based on convolutional neural network is supported by massive data and powerful computing power.With its excellent detection performance,it has become the most concerned research direction in this field.Based on the convolutional neural network,this paper conducts related research on aircraft detection in remote sensing images.The main contents and innovations are as follows:(1)Make MA(Military Aircraft)data set containing different types of aircraft.Based on the convolutional neural network operation mechanism,based on the principle of diverse,balanced,and sufficient image data,3288 satellite remote sensing airport images taken at different periods,different regions,and under different imaging conditions were selected.The production completes five types of fighters,transporters,bombers,tankers,and early warning aircraft,and a total of 10 types of aircraft data sets.(2)A shadow comprehensive processing algorithm based on double-threshold sampling reduction is proposed.Carefully analyze the characteristics of aircraft shadows in remote sensing images,complete the detection of aircraft shadows in Lab color space,and the pixels in the non-shadow area are sampled to restore the shaded area's hue and texture.The experimental comparison shows that the algorithm proposed in this paper can better eliminate the shadow of the aircraft,and the shadow area is smooth and natural to the non-shadow area.(3)A K-Means algorithm based on sample distribution is proposed.According to the statistical characteristics of the sample distribution of the data set,the algorithm adopts the idea of zoning clustering to make the distribution of the data set samples and the clustering value consistent,and fully consider the problem of missed detection of small aircraft during the clustering process.Experiments show that,compared with the traditional K-Means algorithm,the clustering value generated by the algorithm in this paper sets the YOLOv3 anchor parameters,which will effectively improve the detection accuracy.(4)Verify the feasibility of detecting aircraft in remote sensing images based on convolutional neural networks.In this paper,the MA data set was used to train and test Faster R-CNN and YOLOv3,and the experimental results were analyzed qualitatively and quantitatively.Experiments show that both algorithms can detect aircraft.Among them,Faster R-CNN has a slightly higher detection accuracy but a slower speed,and YOLOv3 has a slightly lower detection accuracy but a faster speed.(5)A target detection algorithm(YF R-CNN)based on cross-platform feature fusion is proposed.The core idea of the algorithm is "separate training and joint detection".The key is to integrate the detection feature maps of Faster R-CNN and YOLOv3 across platforms to effectively alleviate the problems of missed and misdetected aircraft.Experiments show that compared with Faster R-CNN and YOLOv3,the algorithm mAP(mean average precision)is improved by 3.1% and 3.7%,respectively.