基于分割块的图像语义分割方法

Image semantic segmentation method based on segmentation block

针对复杂室外环境下,传统语义分割模型无法准确描述对象轮廓的问题,提出了采用结构森林法生成边缘概率,运用分水岭算法将边缘概率转化成初始割块。为避免过分分割,利用超度量轮廓图算法选取适当阈值生成分割块以获取更准确的轮廓信息,通过随机森林训练分割块,得到语义分割结果。实验结果表明:在处理复杂的语义分割任务时,基于分割块的方法在精度、鲁棒性和速率方面均具有良好表现。

Aiming at problem that conventional semantic segmentation models cannot describe object contour accurately under complex circumstance,a new image semantic segmentation method based on segmented block is proposed.Structural forest method is applied to generate contour probability.And the method of watershed is used to transform to initial block of image segmentation. To avoid over-segmentation,ultrametric contour map（UCM）algorithm is performed to select appropriate threshold to generate segmentation block,so as to obtain more accurate contour information. By random forest,train block of image segmentation,obtain semantic segmentation result.Experimental results demonstrate that the proposed method based on segmentation block is superior to traditional methods on precision,robustness and rate while handling complex semantic segmentation task.