基于联合注意力机制的GH4169合金晶粒分割算法研究

Research on the Segmentation Algorithm for GH4169 Alloy Grains Based on a Joint Attention Mechanism

摘要:

针对高温合金晶粒形貌复杂、尺寸差异大、边界模糊且晶粒紧密连接导致分割难题，本文提出一种融合通道与空间信息的联合抑制注意力机制的晶粒度分割网络。该网络结合Swin Transformer的全局建模与CNN的局部细节恢复能力，并在解码器中引入融合通道与空间信息的联合抑制注意力机制，有效抑制噪声和纹理干扰，提升特征筛选与泛化能力，同时强化浅层与深层特征融合，显著改善晶粒边界连贯性。实验结果表明，所提算法在自制金相组织数据集上IoU、F1-score分别可达到67.34%和78.62%，各项指标均优于主流高温合金中的晶粒分割算法。

Abstract:

To tackle the segmentation challenges arising from the complex morphology, significant size variations, blurred boundaries, and tight interconnections of superalloy grains, this paper presents a grain segmentation network integrated with a joint suppression attention mechanism that fuses channel and spatial information. The network combines the global modeling capacity of Swin Transformer and the local detail restoration capability of CNN, and embeds the aforementioned joint suppression attention mechanism, which integrates channel and spatial information, into the decoder. This mechanism effectively suppresses noise and texture interference, enhances the abilities of feature screening and generalization, and reinforces the fusion of shallow and deep features, thereby markedly improving the continuity of grain boundaries. Experimental results demonstrate that the proposed algorithm achieves 67.34% and 78.62% in terms of IoU and F1-score, respectively, on the self-constructed metallographic dataset, with all metrics outperforming those of mainstream grain segmentation algorithms for superalloys.