一种基于目标检测的偏斜图像校准方法

Calibration Method for Inclined Image Based on Object Detection

本文提出了一种新颖的的偏斜图像校准方法,致力于对大规模物联网场景下的偏斜图像进行自动校准.本算法以特征金字塔网络作为基础,提出AFPN网络增加网络特征图的语义信息.由于传统目标检测的方法对非水平目标的匹配会并不能获得目标偏斜角度的信息,因此本文增加偏斜感兴趣区域变换.偏斜感兴趣区域变换中包括了三个阶段,分别为偏斜感兴趣区域学习机,偏斜位置敏感感兴趣区域校准以及偏斜感兴趣区域计算.最后通过对边界框的参数引入direction损失得到边界框偏斜的角度并加权平均得到图像偏斜角度,从而能够对偏斜图像进行校准.

In order to calibrate the angle of the inclined camera in the large-scale IoT scenario,we proposed a novel method that can calibrate the angle of the image.At first,in order to increase the semantic information,we proposed the AFPN network which is based on feature pyramid network.Due to the traditional object detection method can not obtain the angle of target in the image,we then proposed Inclined Rol Transformer to address these problem.The Inclined Rol Transformer consist of three methods which are Inclined RoI(Inclined Rol) learner,Inclined Position Sensitive Rol Align(IPS-RoI-Allign) and Inclined Rol computation.At last,we can acquire the image inclined angle through introducing the direction loss and calculating the weighted average of the inclined angle of bounding box,and then calibrate the inclined camera by using the image inclined angle.