Research On Object Detection In Aerial Optical Remote Sensing Images Based On Regional Proposal

Object detection and recognition on aerial remote sensing images is one of the key technologies in the field of aerial remote sensing,and is widely used in many scenarios such as topographic mapping,resource census and military reconnaissance.The visible light band is the most commonly used working band in traditional aerial reconnaissance photography and aerial mapping photography.The aerial visible band remote sensing images have taken the advantages of high resolution and rich image information,but they are easily affected by factors such as sunlight and cloud interference.Since airborne cameras are located above the object and are far away from the object when shooting the ground,aerial images often show characteristics such as relatively small object occupation,arbitrary object orientation,and complex image backgrounds.At the same time,the ground radiation detected by the camera will be affected by aerosol particles when it passes through the atmosphere,resulting in scattering and absorption.This presents difficulties and challenges for the research of object detection and recognition technology for aerial visible band remote sensing images with high accuracy and good real-time performance.On the background of the visible band remote sensing images taken by high-altitude long-range strabismus aerial cameras,based on the theoretical and technical research of computer vision and image processing,we explored a robust and reliable object autonomous detection and recognition algorithm.Selecting object detection and recognition technology based on region proposal for systematic research,aiming to modify the traditional aerial image object detection and recognition technology.In the mean time,in order to deal with the image degradation caused by the shooting conditions of visible light aerial images,we studied the image dehazing method.This academic dissertation contains the following research contents:First,in order to solve the problem that the visible band remote sensing image is susceptible to cloud and fog interference under actual aerial photography conditions,a feature of image dehazing based on atmospheric transmission degradation model and dark channel prior is proposed.This paper analyzes the degradation mechanism of aerial optical images,the model of atmospheric transmission degradation,and the mathematical basis of the dark primary color prior theory.For the aerial image disturbed by cloud and fog,the traditional dark primary color prior defogging model is extended,and the atmospheric transmission degradation model is introduced,which effectively improves the defogging effect of the aerial image.Secondly,a fast object detection method based on directional area search is proposed.The reason for the low recall rate of Edge Boxes algorithm when detecting objects on aerial images is analyzed,which makes it add a two-step learning SVM linear model after calculating image edge information.The number of candidate regions generated can be effectively constrained efficiently,and the accuracy of region proposals can be improved to obtain preliminary candidate regions.Select the region with higher confidence to apply Radon transform and calculate the variance of the line integral after the transformation to estimate the main orientation of the object in the candidate region.By applying a region boundary fine-tuning algorithm based on superpixel segmentation,after obtaining the pixel block distribution map,the orientation of the bounding boxes of the candidate areas are adjust and similar pixel blocks are merged,so that the main orientation of the candidate area is consistent with the main orientation of the object,and has a high intersection-over-union ratio.The Non-Maximum-Suppression retains candidate regions with high confidence.This method carries the main orientation information of the object and effectively improves the accuracy of detection.Thirdly,a method of object recognition based on candidate regions is proposed.Aiming at the characteristics of the object in the aerial visible band remote sensing image,which are susceptible to noise,blur,and light intensity changes,the robust local LSK(Local Steering Kernel)feature is improved and adopted;aiming at the limitation that the feature is weak in the complex background,the extracted features are integrated into the Vector of Locally Aggregated Descriptors(VLAD)model.By K-means clustering and dictionary coding,the expression ability of feature descriptors is enhanced.A training and classification method of SVM(support vector machine)based on directed acyclic graph is introduced to successfully solve the problem of sample imbalanced caused when training classifying multiple objects,and improve the accuracy of candidate region classification.The results show that the method can successfully achieve accurate recognition of various categories of objects,has good real-time performance,and has good practical effects.Finally,a convolutional neural network algorithm based on fast rotation region proposal is studied.This method is based on the Faster-RCNN model and analyzes the hierarchical structure of the convolutional neural network.Rotating the RPN network makes the generated RoI(region of interest)directional;In the RoI pooling layer,the feature map of the rotating pooling unit is traversed to achieve a better pooling effect;the predicted angle is estimated as a simplified parameter so that the regression can be realized by the bias of object angle.The results show that the advanced convolutional neural network model has improved the recall,accuracy,and average precision in the object detection and recognition scenes of aerial images.Various algorithms proposed in this paper are developed and debugged in a computer environment using Matlab,C / C ++ or Python,and a large number of simulation experiments are performed using a large number of aerial images.Algorithm reliability and robustness are the two main contents of the experimental research in this paper,and the expected experimental results have been achieved in all experiments.