改进的Mask R-CNN算法在人额部区域实例分割任务的应用研究

Improved Mask R-CNN Algorithm in Human Forehead Region Instance Segmentation Task

针对遮盖区域卷积神经网络(Mask Regional Convolutional neural network,Mask R-CNN)在人额部区域分割任务中丢失部分目标的问题,本文改进了Mask R-CNN算法原有的特征金字塔网络(Feature Pyramid Networks,FPN)结构。为了更好地利用图像中反映出的特征信息,首先将原Mask R-CNN中的高维特征信息进行融合,其次,进行ROI Align操作生成人额部的Mask;最后,仿照COCO数据集,从“LIPCIHPinstance-level_human_parsing”数据集中选取带有人脸额部区域的随机场景照片,自建人额部数据集。实验结果表明改进后的FPN网络模型有着更好的目标分割能力,实验效果更好。

Aiming at the problem that Mask regional Convolutional Neural Network(Mask R-CNN)loses some targets in human frontal region segmentation task,this paper improves the original feature pyramid networks(FPN)structure of Mask R-CNN algorithm.In order to make better use of the feature information reflected in the image,firstly,the high-dimensional feature information in the original Mask R-CNN is fused;Secondly,ROI align operation was performed to generate a mask of the adult forehead;Finally,following the coco dataset,from"lipihp instance level"_human_In the"parsing"dataset,select random scene photos with face frontal area to build a human frontal dataset.The experimental results show that the improved FPN network model has better target segmentation ability and better experimental effect.