A Study On Object Detection For Optical Imaging Based On Convolutional Neural Network

Remote sensing technology is a technique that acquires relevant object data or information without any physical contact.Remote sensing technology is widely used in many fields such as military reconnaissance,environmental improvement,and society stability.In recent years,with the rapid development of the aerospace industry,the amount of images obtained by remote sensing satellite has exploded exponentially.Therefore,how to quickly interpret the important information from the massive data becomes especially critical in this scenario.However,object detection of optical remote sensing images is still in the transitional stage between computer-based automatic judgment and manual judgment at present.There are many problems during this stage,such as low efficiency,low timeliness,low intelligence and high cost,leading to the failure of making good use of the massive data resources.The traditional object detection algorithm of remote sensing images can be divided into three steps:region selection,artificial feature extraction and classifier.However,the region selection strategy based on sliding window is not targeted,which can result in several problems such as high time complexity and excessive redundancy of candidate regions.Additionally,the choices of artificial design features are limited and the generalization ability is unsatisfactory.Thus,it is not applicable to massive data.In order to solve the above problems,this thesis establishes a research goal to improve the accuracy of optical remote sensing image object detection based on convolutional neural network.Specifically,objects and backgrounds are difficult to distinguish in the optical remote sensing image,especially when object are dense and small.The thesis proposes an improved two-stage object detection model based on Faster RCNN to solve this problem,which focuses on improving the feature extraction network.Firstly,in order to extract the deeper high-level semantic features,avoid gradient disappearing and maintain a certain speed,the thesis adopts the small residual network ResNet50 instead of the VGG-16 network to upgrade feature extraction backbone network.Secondly,due to the large sampling caused by the deep network,the features of the small objects will not have discriminative power,leading to missed detection and classification errors.The thesis introduces the feature pyramid network to enhance the feature representation,which can combine the low-level combined information and the high-level semantic information.At the same time,generating candidate frames and feature locations are dispersed to each layer of the feature pyramid,which greatly improves the feature mapping resolution of the small objects and attains better detection performance.Thirdly,in order to solve the problem that objects are difficult to distinguish with backgrounds in optical remote sensing image,the thesis adopts a deformable convolution instead of the traditional square convolution,so that the convolution areas are concentrated on the various objects in the network model as much as possible.In detail,the last block of the last three parts of the ResNet50 structure are changed to variable convolution,which reduces background interference.Fourthly,in order to solve the IoU threshold setting on positive and negative sample and further prove the validity of deformable convolution,the thesis introduces the cascade network to improve network performance.Finally,experimental results on the DOTA dataset prove that the proposed two-stage object detection model has achieved satisfactory results in accuracy and has obvious improvement compared with the previous detection models,which indicates that the study in this thesis has practical value.