关键词: Mask RCNN模型, 感兴趣区域池化, Otsu阈值分割, 边缘梯度插值, 双线性插值

Abstract: 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.

Key words: Mask RCNN model, region of interest pooling, Otsu threshold segmentation, edge gradient interpolation, bilinear interpolation