Research on Image Semantic Segmentation Algorithm Based on Fully Convolutional HED-CRF

摘要

Aiming at the problem that the local detail is not accurate enough in 3D scene reconstruction in complex scenes, an image semantic segmentation algorithm based on Holistically Nested Edge Detection-Conditional Random Field(HED) is proposed. First, a Fully Convolutional Networks (FCN) is trained based on Visual Geometry Group Net (VGGNet). The two-stage training method is used to realize the segmentation of the foreground and the background, and the classification image is generated to predict the semantics;Then the edge information is refined by HED algorithm. Finally, the semantic prediction results are optimized by conditional random field. The full convolution network and conditional random field can be integrated into an end-to-end system to achieve accurate segmentation of image semantics. The experimental results show that compared with the traditional machine learning method, the accuracy of this method is improved by 2.9%, and the segmentation effect is better.