Research On Detection Of Aircraft Target Based On Convolutional Neural Network

With the continuous progress of aviation exploration technology and computer vision technology,remote sensing image aircraft target detection is more and more widely used,but with the gradual improvement of optical remote sensing image resolution,its complexity and image scale are also increasing rapidly.For remote sensing image aircraft target detection,there are several main difficulties: in the first remote sensing image aircraft target detection,the size of remote sensing image is larger,but the aircraft has not only smaller fighter aircraft,coach aircraft,but also larger size early warning aircraft,refueling aircraft,passenger aircraft,aircraft target size diversification.Second,the color of many aircraft targets is close to the airport background color,so it is very important to extract shallow semantic information containing detailed texture information.The tensile deformation of the third remote sensing image will have a great influence on the target detection.Because the traditional algorithms need more prior experience to design and have poor adaptability,this paper constructs a multi-scale target detection model based on deep convolution deep network to improve the detection effect of aircraft target in remote sensing image.The main improvement work is as follows:First of all,in order to improve the performance of multi-scale target detection and to separate aircraft targets from similar background,the original feature extraction network is improved.Resnet50+FPN feature extraction network combines high-level semantic features and shallow semantic features through top-down pathways and lateral connections,the feature extraction network based on Resnet50 structure can be used to deepen the network to obtain higher detection accuracy.On the basis of the Resnet50 FPN structure,using the idea of DenseNet,the structure of feature fusion is added in the C1 to C5 layer and P5 to P2 layer of the original feature extraction network,so that the shallow semantic features and the high-level semantic features extracted from the Resnet50 are more fully utilized in the later network.The second,there are large passenger aircraft and smaller military aircraft in remote sensing images.although the MaskRcnn network can achieve semantic segmentation,its semantic segmentation effect is not good in remote sensing images.for this reason,this paper replaces some convolutions with empty convolutions,and then uses the structure of multi-scale feature fusion to better complete the semantic segmentation task.The third,because remote sensing image and natural image have great difference,and MaskRcnn model is proposed for natural image,the design of its anchor frame is more in line with the characteristics of natural images.According to the types of aircraft contained in the data set,this paper designs a set of anchor frame which is more suitable for in remote sensing images according to the scale characteristics of different aircraft.The above method is experimentally verified on the AIRImage data set.when the IoU threshold is 0.5,the improved model achieves 95.8% detection accuracy.Compared with the previous deep learning model,it has better multi-scale target detection effect.Compared with the original MaskRcnn network,the improved network has better detection effect of small object and deformed object,faster convergence,and better semantic segmentation effect on large target of remote sensing image.