Research On Object Detection Based On Local Features For Optical Remote Sensing Imagery

Object detection is an important research direct ion of machine vision and digital image processing technology.It mainly includes feature extraction,feature description and feature matching.At the same time,image feature extraction,description and matching technology is the basis of image mosaicking,target tracking,motion analysis,object recognition and other research fields.The requirements for these applications are different which include: processing time(on-line,off-line,or real-time),robust to occlusions,invarian t to rotations,and optimal detection even viewpoint changes.Nevertheless,the optical remote sensing imagery still incorporates with several limitations of object detection.In accordance with these limitations,this thesis reports the assessment of local features and their improvements.The state-of-the-art feature extraction methods are appropriate according to the categories and structures of the objects to be detected.Similarly,feature matching is the core stage for several applications of computer vision.In object detection,the processing of optical rem ote sensing images becomes critical due to the complex environment,large size,occlusions and color variations.The training process is time-consuming and a large amount of memory is required to store the training images.Numerous traditional state-of-the-art approaches are suffering from problematic high computational time.Meanwhile,Bounding Box Regression(BBR)has played a vital role in object detection.The training through low resolution images yields several unwanted bounding boxes due to arbitrary orientations and complex appearances of objects.Similarly,region proposals have a great impact on object detection and numerous well-known methods(e.g.Selective Search(SS),etc.)strongly attracted the researchers.These region proposals have overcome the limitations of earlier state-of-the-art algorithms such as processing speed,accuracy and computational complexity.In general,these methods are incorporating the multiple objects which reduce the accuracy to extract regions for a single object.The processing of satellite imagery is dependent upon the quality of imagery.Due to low resolution,it is difficult to extract accurate information u nder shadow regions.In accordance with these limitations of optical remote sensing imagery,we have consider ed the proposed solutions as follows:1)Based on distinctive computations of each feature extraction method,different types of images are selected to evaluate the performance of feature extraction methods such as;BRISK,SIFT,SURF,FAST,HOG and LBP.We ha ve studied the combination of SURF with FAST and BRISK individually.We have executed feature matching for newly developed methods in order to provide an optimal solution.Furthermore,RANSAC and MSAC were utilized to eliminate the outliers to get optimal feature matches.2)An efficient region proposal approach has been proposed in comparison with SS region proposal method.The framework is organized into two key steps.The first step is based on extracting region proposals using the Cascade system.The second step is based on the classification of extracted region proposals which is performed by transfer learning using Convolutional Neural Networks(CNN)and AlexNet architecture is utilized for transfer learning.In the meantime,the features are extracted by FAST-SURF combination and SVM is used for classification of extracted region proposals.The comparison between AlexNet and traditional combined features has briefly discussed.We have also proposed a scene-based object detection method.Initially,the particular scene has been classified among multiple scenes.Subsequently,the objects have been detected from that classified scene.The scene classification has performed by transfer learning using AlexNet and object detection is executed by Faster-RCNN.The training and test images have been developed from DOTA dataset and Google earth.3)Firstly,to perceive a better solution to enhance the accuracy for object detection,the post-processing of bounding boxes(BBs)has been evaluated and discussed for the object detection application.Our proposed method is divided into two stages;the first stage is based on thresholding of BBs with respect to the confidence values and the second stage is based on the area-based BB regression(BBR).In BBR,the area of each BB was estimated then the oversized and undersized BBs were removed with respect to the size of objects which are being detected.The widely known region-based approaches RCNN,Fast-RCNN and Faster-RCNN are used for evaluation and comparative analysis validates the proposed framework.Secondly,we have proposed a new training methodology which is based on compressed and down-scaled images,is implemented to reduce the training time.To compare performance,we have trained the RCNN,Fast-RCNN,Faster-RCNN and Cascade detectors by using three types of training image sets and several experiments have been performed on extended VEDAI dataset.4)For the purpose of vehicle detection under shadow regions,we have used HOG for local features extraction,SVM is used for classification.Shadow images have been scrutinized and found very complex for detection as observed very low detection rates therefore our dedication is towards the enhancement of detection rate under shadow regions by implementing appropriate preprocess ing.Vehicles are precisely detected under non-shadow regions with high detection rate than shadow regions.