基于可变向卷积网络的语义分割算法

随着深度学习方法的不断发展,基于深度卷积网络特征的语义分割已经成为自动驾驶、室内导航、遥感制图等领域视觉感知应用的一项重要技术。然而对于多样性变化背景中的目标图像,现有基于局部上下文卷积特征的语义分割方法仍然存在分类精度低的问题。为此,提出了基于可变向卷积网络的语义分割算法。首先,在特征图每一个像素点上预测对象主要观测方向,然后在这个主要的方向上通过卷积运算来预测对象的类别。考虑到对象的尺度的变化,算法在预测像素点类别时,使用空洞卷积在多个尺度下进行预测。相比于一般的语义分割网络,算法在多个方向选择性地利用显著的语义特征,融入更可分的上下文,提高了网络的识别能力。算法在PASCAL VOC2012公开数据集上取得了更优的结果。

A Segmentation Method Based on Significant Semantic Features of Deep Convolution Network

With the continuous development of deep learning methods,semantic segmentation based on deep convolutional network features has become an important technology for visual perception applications in the fields of autopilot,indoor navigation and remote sensing mapping.However,for the target image in the background of diversity change,the existing semantic segmenta?tion method based on local context convolution features still has the problem of low classification accuracy.So a semantic segmenta?tion algorithm based on variable convolutional networks is proposed.First,the main observation direction of the object is predicted at each pixel of the feature map,and then the convolution operation is used to predict the category of the object in this main direc?tion.Considering the change in the scale of the object,the algorithm uses the atrous convolution to predict at multiple scales when predicting the pixel class.Compared with the general semantic segmentation network,the algorithm selectively utilizes significant semantic features in multiple directions,integrates into more divisible contexts,and improves the recognition ability of the network.The algorithm achieves better results to other data in the PASCAL VOC2012 public dataset.