A Research On Object Detection In Remote Sensing Images Based On The Combination Of Eye Tracking And Computer Vision

Remote sensing(RS)object detection and recognition has always been a hot topic in the field of remote sensing image processing and pattern recognition.The localization of objects of interest has very important practical significance both in civilian and the military field.Facing the staggering growth of geometric data on account of the rapid development of geostationary observation technology,the traditional way to obtain the information of the locations of targets information relying solely on artificial visual interpretation has been far from being able to meet the needs of modern society for efficient information,due to its low efficiency,highly subjective,high cost,long information acquisition cycle and other defects.And there are also the problems of poor accuracy and other issues exist when simply rely on machine vision method.Therefore,how to combine the advantages of both human and machine to automatically extract and identify sensitive targets from the massive remote sensing data quickly and accurately has became an problems need to be solved.Eye movement measurement provides a method of capturing cognitive processes at high resolution,In this paper,we focus on the object detection in remote sensing images based on the combination of eye movement and computer vision.This paper base on research on the practical application point of view,combining both the eye movement and image information,and our main achievements are as follows:1.As a result of the complex background in most of the RS images and the strong interference caused by it,the algorithms use only image information can hardly locate the target fast and accurately.This paper introduce a kind of method of saliency analysis.The introduction of Gaussian mixture model with the weights learning from different gaze features for each component can obtain fixation-oriented Gaussian map,which would locate the targets more specifically due to the absence of image features.Then the fixation-oriented Gaussian map was further optimized by the single layer cellular automata with the image information to obtain the attention related saliency map.The proposed saliency map combine the cognitive and image information,reduce the background interference.Experiments have showed that the proposed algorithm can be used to calibrate the salient target in the presence of complex background and strong interference.2.In order to solve the problem that the number of candidate bounding boxes in the target area is large and the following high cost as well as low efficiency in detection period,this paper proposed a single object detection method based on the proposed saliency map.It integrate fringe restraint into both the instruction to guide the super-pixel fusion and further extraction of the bounding-boxes and the bounding-box scoring process to generate a very small set of bounding-box proposals.Also,a bounding-box scoring criteria to determine the final candidate box is proposed with the combination of eye movement,saliency and edge information.The proposed method introduce the semantic information represented by eye movement into object detection,makes the use of human visual mechanism to rapidly screen suspicious areas and improve both the accuracy and efficient impressively.Experiments show that the detection results of the general targets can reach comparable results with the state-of-the-art methods.Moreover,this method is more prominent on the small scale target and the effect increased more than 50%.3.Aiming at the problem that saliency detection shows less efficiency on multi-target images and the existing algorithms are lack of common criteria that can generalize them to common tasks.In order to meet the needs of the RS image object detection in real world,a computational framework for candidate box selection and the following bounding-box regression is proposed.In order to improve the final result,the bounding-box criteria is adjusted for multi-target detection by extracting corresponding bounding-boxes from every fixation point,and the bounding-box proposal is selected based on the combination of fixation and IOU score.Moreover,the final bounding-boxes will be optimized by edge regression using deep learning.Experiments show that the proposed method shows competitive detection results on remote sensing images,and successfully reduce the rate of false positive and false negative.