基于空洞卷积的人体实例分割算法

Person Instance Segmentation Algorithm Based on Dilated Convolution

针对人体实例分割任务中存在着姿态多样和背景复杂的问题,提出了一种高精度的实例分割算法。利用Mask R-CNN算法特征融合过程中的细节信息,改善人体分割任务中边缘分割不精确的问题,提高人体分割精度。改进了特征金字塔的特征融合过程,将原有自顶向下的路径改为自底向上,以保留浅层特征图中更多的空间位置信息,并且在特征融合过程中加入多尺度空洞卷积,在增大特征图感受野同时保持分辨率不变,可以避免降采样过程中特征丢失;使用COCO数据集和网络平台,建立新的人体图像数据集。最后将本算法与Mask R-CNN算法做对比,在IoU分别为0.7、0.8和0.9时,准确率提高了0.26,0.41和0.59。实验结果表明,算法在新的人体图像数据集可以得到更精确的结果。

Aiming at the problem of multiple pose and complex background in person instance segmentation,a high-precision instance segmentation algorithm is proposed.By making use of the detailed information in the process of feature fusion of Mask R-CNN,the problem of inaccurat edge segmentation in person instance segmentation is improved,and the accuracy of person instance segmentation is improved.The feature fusion process of feature pyramid is improved.The original top-down path is changed to bottom-up path to retain more spatial details in the shallow feature map,and multi-scale dilated convolution is added in the feature fusion process to increase the receptive field of the feature map while keeping the resolution feature map unchanged,which can avoid information loss caused by down sampled.A new person image dataset is established from COCO dataset and network platform.Finally,compared with the Mask RCNN,the proposed algorithm improves the accuracy by 0.26,0.41 and 0.59 when the IOU is 0.7,0.8 and 0.9 respectively.Experimental results show that algorithm can get more accurate person segmentation effect on the new person image dataset.