一种边缘梯度插值的感兴趣区域池化算法

A Region of Interest Pooling Algorithm forEdge Gradient Interpolation

针对现有主流的目标检测算法存在检测精确率低、图像边缘区域分割不全等问题,提出一种基于Mask RCNN模型的感兴趣区域池化算法.首先,通过Otsu阈值分割法将感兴趣区域特征图划分为边缘区域和非边缘区域;其次,对边缘区域使用边缘梯度插值算法进行插值,对非边缘区域使用双线性插值算法进行插值,从而将离散的特征图映射到一个连续空间中;再次,将插值后的特征图均匀分割成k×k个单元;最后,对每个单元利用二重积分求均值以完成池化操作.对比实验结果表明,该算法基于Mask RCNN模型在数据集COCO(2014)上比现有算法的检测精确率有一定提升,对图像边缘区域的细节分割效果较好.

Aiming at the problems that the existing mainstream target detection algorithms had low detection accuracy and incomplete segmentation in the image edge regions,we proposed a region of interest pooling algorithm based on Mask RCNN model.Firstly,the feature maps of the regions of interest were divided into edge regions and non-edge regions by the Otsu threshold segmentation method.Secondly,the edge gradient interpolation algorithm was used to interpolate for the edge regions,and the bilinear interpolation algorithm was used to interpolate for the non-edge regions so that the discrete feature map was mapped into a continuous space.Thirdly,the interpolated feature maps were evenly divided into k×k units.Finally,the double integral was used to calculate the average value of each unit to complete the pooling operation.The comparative experimental results show that the proposed algorithm,based on the Mask RCNN model,has a certain improvement in detection accuracy compared with existing algorithms on COCO(2014)dataset,and has a good segmentation effect on the details of the image edge regions.