Lightweight cross scale feature fusion for microstructural inclusion detection in aluminum alloys using deep learning

Tianlong Wang, Ying Li, Yushi Ding, Yunlong Hao, Zhenwei Liu, Chunsheng Zhuang, Wei Zhang,
Lightweight cross scale feature fusion for microstructural inclusion detection in aluminum alloys using deep learning,
Materials Today Communications,
Volume 49,
2025,
113938,
ISSN 2352-4928,
https://doi.org/10.1016/j.mtcomm.2025.113938.
(https://www.sciencedirect.com/science/article/pii/S235249282502450X)
Abstract: Aluminum alloys are indispensable to mechanical manufacturing, aerospace engineering, and energy systems. However, the supply of high performance materials remains limited owing to the structural complexity and typological diversity of microstructural inclusions. This constraint necessitates sophisticated detection models capable of precisely identifying complex alloy microstructures. Herein, we present a deep learning based fully convolutional cross scale feature fusion framework for inclusion detection in aluminum alloys. A lightweight fully convolutional neural network, developed while preserving the original architecture, establishes an optimal balance between global information extraction and parameter efficiency. The incorporation of a learnable local saliency kernel module enhances structural feature discrimination, whilst a cross scale selective fusion module facilitates precise multi scale feature integration, thereby yielding considerable improvements in sensitivity and robustness for multi scale target perception. Experimental results demonstrate that the proposed model surpasses mainstream detection algorithms, achieving 91.3% mAP@0.5 and 61.1% mAP@[0.5:0.95], while substantially reducing the parameter scale. This work provides a robust technical foundation for machine vision based intelligent inclusion detection and removal in aluminum alloys, and promotes the application of deep learning in materials science and microstructural analysis. The code is available at https://github.com/wtl985/CSF-YOLO.
Keywords: Deep learning; Microstructure analysis; Alloy microstructures; Feature fusion; Fully convolutional