曹富强, 王明泉, 张俊生, 邵亚璐, 张雪洋. 基于深度学习的铸件X射线图像分割研究[J]. 应用光学, 2021, 42(6): 1025-1033. DOI: 10.5768/JAO202142.0602003
引用本文: 曹富强, 王明泉, 张俊生, 邵亚璐, 张雪洋. 基于深度学习的铸件X射线图像分割研究[J]. 应用光学, 2021, 42(6): 1025-1033. DOI: 10.5768/JAO202142.0602003
CAO Fuqiang, WANG Mingquan, ZHANG Junsheng, SHAO Yalu, ZHANG Xueyang. Casting X-ray image segmentation based on deep learning[J]. Journal of Applied Optics, 2021, 42(6): 1025-1033. DOI: 10.5768/JAO202142.0602003
Citation: CAO Fuqiang, WANG Mingquan, ZHANG Junsheng, SHAO Yalu, ZHANG Xueyang. Casting X-ray image segmentation based on deep learning[J]. Journal of Applied Optics, 2021, 42(6): 1025-1033. DOI: 10.5768/JAO202142.0602003

基于深度学习的铸件X射线图像分割研究

Casting X-ray image segmentation based on deep learning

  • 摘要: 针对当前图像分割算法在实现工业铸件内部缺陷分割上精度低且算法不够轻量化的问题,提出一种基于改进DeepLabv3+的工业铸件内部缺陷检测算法Effi-DeepLab。该方法采用EfficientNet中的MBConv来代替原有的Xception模块进行特征提取,使特征提取网络更加高效与轻量化;针对工业铸件内部缺陷尺寸小的问题,重新设计空洞空间金字塔池化(ASPP)层中空洞卷积的扩张率,使得卷积块对小目标具有更高的鲁棒性;在解码端充分利用特征提取阶段的低阶语义信息进行多尺度特征融合,以提高小目标缺陷分割的精度。实验结果表明,在本文使用的汽车轮毂内部缺陷图像数据集中,Effi-DeepLab模型对缺陷的分割准确率和平均交并比(mIoU)分别为93.58%和89.39%,相比DeepLabv3+分别提升了2.65%和2.24%,具有更好的分割效果;此外,还通过实验验证了本文提出算法具有良好的泛化性。

     

    Abstract: Aiming at the problems that the current image segmentation algorithm has low accuracy in realizing the internal defect segmentation of industrial castings and the algorithm is not lightweight enough, an improved industrial casting internal defect detection algorithm Effi-DeepLab based on DeepLabv3+ was proposed. This method used MBConv in EfficientNet to replace the original Xception module for feature extraction, making the feature extraction network more efficient and lighter. Aiming at the problem of small internal defects in industrial castings, the expansion rate of the hollow convolution in the atrous spatial pyramid pooling (ASPP) layer was redesigned. The convolution block was more robust to small targets; the low-order semantic information in the feature extraction stage was fully utilized at the decoding end to perform multi-scale feature fusion to improve the accuracy of small target defect segmentation. The experimental results show that the segmentation accuracy and mIoU of the Effi-DeepLab model in the internal defect image data set of the automobile wheel are 93.58% and 89.39%, respectively, which are improved by 2.65% and 2.24%, respectively, compared with DeepLabv3+, and has better segmentation effect. In addition, it is experimentally verified that the proposed algorithm has good generalization.

     

/

返回文章
返回