自适应特征融合的高铁轴承金相图像分割方法研究

Adaptive Feature Fusion-Based Segmentation Method for Metallographic Images of Heavy-duty Bearings

摘要:

重载轴承金相图像的精准分割是轴承微观结构评估、服役寿命预测的重要手段。针对重载轴承金相图像中的碳化物组织边缘清晰度低、尺寸差异大与颗粒较小等导致分割精度问题，提出了一种基于自适应多尺度特征融合与注意力机制的金相图像碳化物分割模型ACSA-U-net。通过构建多尺度特征提取网络，在编码器路径的跳跃连接处嵌入了CASAB模块，强化跳跃连接传递的特征质量，突出有效通道与空间区域，抑制无关背景干扰，提升了金相图像中碳化物颗粒的特征信息捕获能力；在解码器路径上的采样过程中设计了自适应特征融合模块AFEM，采用自适应门控机制，实现了多尺度特征的自适应融合与增强，并在其中嵌入CBAM模块，在通道和空间两个维度上对特征图进行自适应优化。基于某型重载轴承金相图像数据集的验证结果表明，所提ACSA-U-net模型在Dice系数、精确率、召回率与交并比均优于传统U-net++及其改进模型，实现金相图像碳化物的稳定分割，为重载轴承的服役性能评价研究提供了支撑。

Abstract:

Accurate segmentation of metallographic images of heavy-duty bearings is an important approach for evaluating their microstructures and predicting service life. Aiming at the segmentation accuracy issues caused by low edge definition, large size variations and small particle sizes of carbide microstructures in heavy-duty bearing metallographic images, an ACSA-U-net model for carbide segmentation in metallographic images is proposed based on adaptive multi-scale feature fusion and attention mechanism. By constructing a multi-scale feature extraction network, the CASAB module is embedded into the skip connections of the encoder path to enhance the quality of features transmitted via skip connections, highlight effective channels and spatial regions, suppress irrelevant background interference, and improve the ability to capture feature information of carbide particles in metallographic images. An Adaptive Feature Enhancement Module (AFEM) is designed in the sampling process of the decoder path, which adopts an adaptive gating mechanism to realize adaptive fusion and enhancement of multi-scale features. The CBAM module is embedded therein to adaptively optimize the feature maps in both channel and spatial dimensions. Validation results based on a metallographic image dataset of a certain type of heavy-duty bearing show that the proposed ACSA-U-net model outperforms the traditional U-net++ and its improved variants in terms of Dice coefficient, precision, recall and intersection over union (IoU). It achieves stable segmentation of carbides in metallographic images and provides support for the research on service performance evaluation of heavy-duty bearings.