融合全景分割的单目深度估计网络

MONOCULAR DEPTH ESTIMATION NETWORK WITH FUSED PANORAMIC SEGMENTATION

针对遮挡和杂乱光线导致的不同区域深度边缘模糊、边界伪影等问题,提出一种结合多任务轻量型卷积神经网络的单目深度图像估计方法。利用全景分割网络来辅助单幅图片的深度估计,选择MobileNetv2作为特征提取网络,解码器端融合以上两类任务进行相似性辅助决策。提出一种多任务融合模块,包括多尺度映射单元和多任务融合单元两部分,利用深度空洞卷积扩大不同感受野,融合多任务来优化深度图像的估计。此外编解码器结构之间添加跳跃连接实现不同层次的知识传递。在NYUdepth-v2数据集上的对比实验结果表明,该方法深度图估计结果更加清晰,并能有效去除深度图中的边界模糊,同时该网络在参数数量上相较大多数估计方法大幅度减少,准确率明显提升。

Aiming at the edge blur and boundary artifacts in different regions caused by occlusion and stray light,this paper proposes a monocular depth image estimation method combined with the multi-task lightweight convolution neural network.It applied the panoramic segmentation network to assist the depth estimation of a single image.MobileNetv2 was selected as the feature extraction network,and the decoder fused the above two kinds of tasks for similarity decision-making.It proposed a multi-scale fusion module including multi-scale mapping unit and multi task fusion unit,which adopted the depth dilated convolution to expand different receptive fields,and fused multi tasks to optimize the depth image estimation.It added jump connection between codec structures to realize knowledge transfer at different levels.The comparative experimental results on the NYUdepth-v2 dataset show that the depth map estimation results of this method are clearer,and this method can effectively remove the boundary ambiguity in the depth map.Meanwhile,the parameter number of this network is greatly reduced compared with most estimation methods,and the accuracy rate is significantly improved.