A Semantic Segmentation Model for Substation Scenes Based on Monocular Depth Estimation
In response to the lack of effective learning of three-dimensional deep geometric information in existing semantic segmentation methods,which leads to low accuracy in object semantic segmentation in complex substation scenes,this article proposes a semantic segmentation model based on monocular depth estimation of substation scenes.This model consists of two parts:DeepLab v3+auxiliary image semantic segmentation model and AdaBins Module monocular depth estimation model.Firstly,The AdaBins Module generates corresponding depth maps based on visible light images,thereby extracting depth information of the target object in three-dimensional space from the images.Secondly,the depth information in the depth map is fused with visible light images as weights using matrix multiplication,and the invalid background pixels in the image are weakened based on the established depth threshold to reduce their impact on the accuracy of target object segmentation in subsequent image segmentation stages.Finally,input the fused image into the DeepLab v3+auxiliary image semantic segmentation model for semantic segmentation.Experiments have shown that compared to the benchmark model,the method proposed in the article can better extract the depth contour features of segmented targets,and the semantic segmentation accuracy is significantly improved.
NETL
NSTL
万方数据
