Aircraft Detection Research In Remote Sensing Images

The continuous advancement in the field of sensor manufacturing has greatly promoted the rapid development of modern remote sensing technology,which has brought about a huge improvement in the quality and quantity of remote sensing images.In the face of massive remote sensing images,it is of great practical significance and application value to research how to extract and utilize effective information quickly and accurately.Object detection in remote sensing images is one of the important directions of remote sensing image application.The automatic interpretation of remote sensing images can be further realized by accurate object detection.As a special strategic object,the aircraft has a certain practical significance for the research on the problem of object detection in remote sensing images.In recent years,the deep learning algorithm has performed well in object detection in natural images,and the detection accuracy is significantly better than the traditional object detection algorithm with artificial designed features.In view of this,the research focuses on the object detection algorithm based on convolutional neural network.According to different practical needs,this research starts from the idea of multi-scale receptive field fusion,and improves the detection algorithms of two different genres respectively,and proposes aircraft object detection models suitable for remote sensing images.The main research work is as follows:(1)A high-precision aircraft detection model based on multi-branch RPN is proposed.Based on the typical two-stage object detection algorithm Faster R-CNN,the network model is improved from three aspects to achieve high-precision aircraft detection: Firstly,replace the backbone network with ResNet to increase the network depth and enhance the feature expression capability,through multi-level feature fusion to obtain different levels of feature information to improve detection performance for small objects;secondly,expand the scale range of anchor to increase the probability that the aircraft object is covered by the preset frame;thirdly,design the multi-branch RPN structure to extend the receptive field of RPN by fusing convolution kernel branches of different sizes,and improve the adaptability of the network to different scale changes.The improved algorithm has precision of 99.55% and recall of 98.26% on the self-built dataset.The detection precision and recall on the public dataset are higher than the similar algorithms and traditional algorithms.The experimental results show that the improved algorithm has obvious improvement effect on detection accuracy.(2)A high-performance aircraft detection model based on multi-branch deformable convolution is proposed.The two-stage detection method takes a long time and is suitable for offline calculation scenarios with high accuracy requirements.It is not suitable for application scenarios with high real-time requirements.Select the faster one-stage detection method SSD as the basic network model,and improve model by adding lightweight feature modules to achieve high-performance aircraft detection.On the one hand,using RFB module to fuse the receptive fields of multiple resolutions,increasing the robustness of the network to aircraft object scale changes.On the other hand,connecting the deformable convolution layer after the dilated convolution layer of each branch of the RFB module.So that the convolution layer is no longer regular equally spaced sampling,but adaptively adjusts the spatial distribution of the sampling position according to the scale and shape of the aircraft object,thereby improving the detection performance of the network.The improved algorithm's detection accuracy on self-built dataset and public dataset is comparable to the two-stage detection algorithm,while detecting a single image takes only 33 ms.The experimental results show that the improved algorithm can significantly improve the detection accuracy while not affecting the speed of the model,and meet the needs of real-time detection to a certain extent.