【作者】 杨帆；

【导师】 宋晴；

【作者基本信息】 北京邮电大学 ， 控制科学与工程， 2022， 硕士

【摘要】 近年来,卫星、无人机等遥感技术得到了快速发展,遥感数据呈现出体量大、来源和类型多、更新快等大数据的特点。海量遥感图像对图像处理技术提出了极大挑战,也带来了新的机遇。本文面向光学遥感图像,研究一套准确且高效的目标检测算法,为救灾应急、交通监管等高层级智能任务提供坚实的基础。本文开展了如下研究工作:首先,针对基于回归的斜框检测器面临的回归目标不连续问题,本文研究了基于极坐标的和基于关键点估计的斜框检测器,通过对斜框建模的改变,避免了边框回归中对斜框角度的预测。实验表明,基于极坐标的检测器,以极角降序排列四个端点,可以最大程度地缩小预测的动态范围;通过独立预测四个绝对极径,增加结果的容错性。基于关键点估计的检测器通过热图预测斜框四个端点的响应,取得了较高的基准精度。其次,本文进一步研究了基于关键点估计的检测算法。针对分类得分与检测质量不一致问题,本文使用端点置信度的平均值量化检测框的定位质量,并与分类得分修正得到检测得分。针对图像切分导致的端点丢失问题,本文额外预测斜框斜外接矩的几何中心,根据预测端点的对角线中点是否足够接近中心点来判断是否将预测框修正。两项策略为算法精度带来了稳定的提升。最后,为了让检测器的推理更加高效,本文研究了从粗略到精细的遥感大图推理策略。通过轻量化的语义分割模型预测图像中是否存在目标框,仅对可能存在目标的子图像进行精细的检测。在最大程度地保留检测精度的情况下,推理速度可提高30%。更多还原

【Abstract】 In recent years,remote sensing technologies such as satellites and UAVs have developed rapidly.Remote sensing data presents the characteristics of big data such as large volume,various sources and types,and fast updating.Massive remote sensing images have brought great challenges to image processing technology but also brought new opportunities.This paper studies an accurate and efficient object detection algorithm for optical remote sensing images,which provides a solid foundation for high-level intelligent tasks such as disaster relief and emergency,traffic supervision,and so on.This paper has carried out the following research work:First,aiming at the discontinuity of regression targets faced by regression-based oriented object detectors,this paper studies detectors based on polar coordinates and key-point estimation.By changing the modeling of the oriented box,the prediction of the angle in the bounding box regression is avoided.Experiments show that the detector based on polar coordinates can minimize the dynamic range of prediction by arranging the four endpoints in descending order of polar angles;by independently predicting the four absolute polar diameters,the fault tolerance of the results increased.The detector based on key-point estimation predicts the response of the four endpoints of the box through the heat map and achieves a high benchmark accuracy.Secondly,this paper further studies the detection algorithm based on key-point estimation.Aiming at the inconsistency between classification score and detection quality,this paper uses the average value of endpoint confidence to quantify the positioning quality of the detection box and corrects it with the classification score to obtain the detection score.Aiming at the problem of endpoint loss caused by image segmentation,this paper additionally predicts the geometric center of the oblique circumscribed moment of the box and judges whether to correct the prediction frame according to whether the midpoint of the diagonal of the predicted endpoint is close to the center point.The two strategies bring stable improvement to the accuracy of the algorithm.Finally,to make the inference of the detector more efficient,this paper studies a large-scale remote sensing inference strategy based on coarse-tofine.A lightweight semantic segmentation model is used to predict whether there is an object box in the image,and only finely detect sub-images where there may be objects.Inference speed can be increased by 30%with maximum preservation of detection accuracy.更多还原