| Remote sensing image analysis based on deep learning technology has attracted much attention in recent years.The detection of vehicle in remote sensing images has broad application prospects in traffic management and control.The existing remote sensing image vehicle target detection algorithm is mainly implemented based on anchor and it is difficult to achieve a balance between detection accuracy and speed.On the one hand,the existence of anchor increases the amount of calculation of the target detection algorithm.On the other hand,the manually set anchor parameters are relatively discrete,which makes it difficult to effectively and completely match the vehicle target in the remote sensing image.Therefore,this thesis introduces the anchor free detection method into the remote sensing image vehicle detection,and on the premise of eliminating the negative impact caused by the anchor,it ensures that the detection model has both high detection accuracy and detection speed.The main work of this thesis is as follows:(1)In view of the small vehicle targets in the remote sensing image data set,and the complex image background information,there are more image distortions,noises,and vehicle-like targets.This thesis proposes the MA-FPN(Multi-Attention Feature Pyramid Network)module for background information filtering and Enhanced vehicle targets.The MA-FPN module integrates a variety of attention mechanisms into the traditional Feature Pyramid Network(FPN)structure,and enhances the vehicle targets in the remote sensing image based on the spatial and channel features of the image,so that the interference information is suppressed and effective information is strengthened.Experiments show that this module has a significant effect on improving the detection accuracy without significantly reducing the model detection rate.(2)Aiming at the problem of inconsistent selection criteria for training samples in the anchor free detection algorithm,and considering the large amount of redundant information in the horizontally labeled rectangular box of the remote sensing image vehicle target,this thesis proposes the MPFA(More Precise Fovea Area)method to determine the training sample selection area of the vehicle target in the remote sensing image.In this thesis,based on the characteristics of the arbitrary direction of the vehicle in the remote sensing image,a new method of selecting the training sample area is proposed.By simultaneously inputting the horizontal rectangular frame of the vehicle and the instance segmentation information during the training process,an irregular area with directional information and capable of representing vehicle distribution is determined,and the pixels within the range will be used as an anchor free detection method.The experimental results show that this method has achieved significant improvement effects on the two data sets in this thesis.(3)Aiming at the shortcomings of the public data set VEDAI(Vehicle Detection in Aerial Imagery),a private data set RSI(Remote Sensing Image)was designed and produced as a supplement to VEDAI.The VEDAI data set has been produced for a long time,the distribution of vehicle targets in the data set image is discrete and the background information of the picture is not rich enough.As a supplement to the public data set VEDAI data set,focusing on the shortcomings of the VEDAI data set,the RSI data set was designed and produced pertinently. |
