基于Deeplabv3工程图形的轻量化分割方法

Lightweight Segmentation Method Based on DeeplabV3 Engineering Graphics

重工装备制造,通过分割工程图纸中所需特定形状,按照图纸比例加工制造。目前,工程图轮廓分割提取大都通过人工实现,其效率大大降低。因此,提出一种基于Deeplabv3的工程图轮廓分割提取方法。该方法能够自动提取工程图中所需零件轮廓并进行下料,大大加快制造效率。Deeplabv3采用级联或并行和多个不同空率的空洞卷积模块实现图像特征的提取,摒弃了传统DeeplabCRF处理,在分割图像方面有着出色的性能。文章对经典Deeplabv3进行改进,提出用轻量化MobileNet作为主干网络,并首次用于工程图分割方面。经过实验分析,该方法在工程图数据集分割方面有着不错的效果。

Heavy engineering equipment manufacturing, by segmenting the specific shape required in the engineering drawings, according to the proportion of the drawings processing manufacturing. At present, most of the engineering drawing contour segmentation and extraction is achieved manually, and its efficiency is greatly reduced. Therefore, a Deeplabv3-based engineering drawing contour segmentation and extraction method is proposed. The method can automatically extract the required part contour in engineering drawings and undergo materialization, which greatly accelerates the manufacturing efficiency. deeplabv3 adopts cascade or parallel and multiple null convolution modules with different null rates to achieve the extraction of image features, which discards the traditional Deeplab CRF processing and has excellent performance in segmenting images. The article improves the classical Deeplabv3 by proposing a lightweight MobileN et as the backbone network, and is used for the first time in engineering image segmentation. After experimental analysis, the method has good results in segmenting engineering graph datasets.