基于图注意力机制超分网络模型的土建场景实例分割

Learning Instance Segmentation of Civil Engineering Scene Based on Graph Attention Mechanism and Super-Resolution Network Model

针对变电站土木建设现场智能识别中的建筑工地施工场景复杂、目标识别分割困难及信息不对称等问题,本文提出了一种基于图注意力机制超分网络模型的智能识别场景解析技术,利用图卷积神经网络和注意力机制网络提取图像目标特征深层次信息,通过双线性插值与反卷积相结合的像素超分辨率技术处理得到清晰的图像目标物边界,实现变电站土建场景实例分割。结果表明,图注意力机制超分网络模型有效解决了航拍图像中目标物边界信息不丰富、实例分割精度差等问题,准确分割变电站场景目标物,目标物边缘界限清晰。

There are a number of problems in the intelligent recognition of substation civil construction site,for instance,complexity of construction scene,difficulty of target recognition and segmentation,and asymmetry of information,etc.To solve the problems above,this paper proposes an intelligent scene recognition and analysis technology based on the graph attention mechanism and super-resolution network model,which uses the graph volume spirit network and attention mechanism network to extract the deep-seated features of target images.Through the pixel super-resolution technology combined with bilinear interpolation and deconvolution,it is feasible to obtain clear the object boundary in images and realize the instance segmentation of civil engineering scene.The results in the paper show that the graph attention mechanism and super-resolution network model effectively solves the problems of insufficient target boundary information and poor case segmentation accuracy in aerial images.What’s more,it can accurately segment the target in substation scene,and the edge boundary of target is clear.