重庆理工大学学报(自然科学) ›› 2023, Vol. 37 ›› Issue (6): 187-195.

• 信息·计算机 • 上一篇    下一篇

交叉门控融合的改进语义分割网络及应用

陈海永,刘新如   

  1. (河北工业大学 人工智能与数据科学学院,天津 300130)
  • 出版日期:2023-07-12 发布日期:2023-07-12
  • 作者简介:陈海永,男,博士,教授,主要从事图像处理、机器视觉和模式识别研究,Email:haiyong.chen@hebut.edu.cn。

An improved semantic segmentation network and its application by using cross-gated fusion

  • Online:2023-07-12 Published:2023-07-12

摘要: 针对太阳能电池表面的断栅、划痕、黑斑等导致的缺陷分割精度差的问题,提出一 种交叉门控融合的改进语义分割网络。使用门控机制选择性地融合网络中的多尺度信息,充分 利用底层细节信息和高层语义信息,增强微小缺陷的特征表示,并结合上下文模块提高获取全 文信息的能力。为了进一步解决太阳能电池缺陷边缘信息弱的问题,引入 PointRend模块对缺 陷边缘的点进行采样,对边缘中不确定的点实行自适应细分策略,实现对缺陷边缘的精细分割。 实验结果表明:所提方法在太阳能 EL组件电池数据集上的 mIoU达到了 65.53%。和现有的语 义分割算法相比,所提方法能够有效细化目标边界,更好地处理微小微弱缺陷。

关键词: 太阳能电池, 缺陷分割, 多尺度特征, 门控融合, 上下文注意力

Abstract: This paper proposes an improved semantic segmentation network based on cross-gated fusion due to poor segmentation accuracy caused by small and weak defects such as broken grids,scratches and black spots on the surface of solar cells.Firstly,to address the problem of small defects in solar cells,a cross-gated fusion module is proposed,which selectively fuses multi-scale information in the network through the gating mechanism.The module makes full use of low-level detailed information and high-level semantic information,enhances the feature representation of small defects,and improves the ability to obtain full-text information by combining the context module.Secondly,to further solve the problem of weak edge information of solar cell defects,the PointRend module is introduced to sample the points on the defect edge,and an adaptive subdivision strategy is implemented for the uncertain points on the edge to realize fine segmentation of the defect edge.Finally,the experimental results show that the mIoU of the proposed method on the data set of solar EL component batteries reaches 65.53%.Compared with the existing semantic segmentation algorithm,the proposed method can effectively refine the target boundary and handle small and weak defects better.

中图分类号: 

  • TP391.4