山东理工大学学报（自然科学版）

为提高图像配准算法的精度和速度，提出一种使用聚焦线性注意力转换器的局部特征匹配算法LoFLAT。LoFLAT由3个主要模块组成：特征提取模块、特征转换模块和匹配模块。使用ResNet和特征金字塔网络提取分层特征；特征转换模块采用聚焦的线性注意力，以利用聚焦映射函数细化注意力分布，并利用深度卷积增强特征多样性；匹配模块通过由粗到精的策略预测准确且鲁棒的匹配。实验证明，本文算法与SIFT算法、SURF算法、ORB算法和改进SIFT算法相比，在纹理差和纹理丰富的区域都能够提取更多、更精确的特征点，且匹配速度更高。

In order to improve the accuracy and speed of the image registration algorithm, this study proposes a local feature matching method named LoFLAT, using a focused linear attention converter. LoFLAT consists of three main modules: a feature extraction module, a feature conversion module and a matching module. The feature extraction module utilizes ResNet and feature pyramid network to extract hierarchical features. The feature converter module refines the attention distribution using focused linear attention and the focused mapping function, and enhances feature diversity through deep convolution. The matching module predicts accurate and robust matching through a coarse-to-fine strategy. Experimental results show that compared with SIFT, SURF, ORB and improved SIFT algorithms, LoFLAT algorithm extracts more and more accurate feature points in both texture-poor and texture-rich areas, while achieving higher matching speed.