CMFNet: A Three‐Stage Feature Matching Network With Geometric Consistency and Attentional Enhancement

ABSTRACT Current feature matching methods typically employ a two‐stage process, consisting of coarse and fine matching. However, the transition from the coarse to the fine stage often lacks an effective intermediate state, leading to abrupt changes in the matching process. This can hinder smooth transitions and precise localization. To address these limitations, this study introduces Coarse‐Mid‐Fine Match Net (CMFNet), a novel three‐stage image feature matching method. CMFNet incorporates an intermediate‐grained matching phase between the coarse and fine stages to facilitate a more gradual and seamless transition. In the proposed method, the intermediate‐grained matching refines the correspondences obtained from the coarse‐grained stage using Adaptive‐random sample consensus (RANSAC). Subsequently, the midtransformer, which integrates sparse self‐attention (SSA) mechanisms with local‐feature‐based cross‐attention, is employed for feature extraction. This approach enhances the feature extraction capabilities and improves the adaptability to various types of image data, thereby boosting overall matching performance. Additionally, a cross‐attention mechanism based on local region features is introduced. The network undergoes fully self‐supervised training, aiming to minimize a match loss that is autonomously generated from the training data using a multi‐scale cross‐entropy method. A series of thorough experiments was carried out on diverse real‐world datasets, including both unaltered and extensively processed images.The results demonstrate that the proposed method outperforms state‐of‐the‐art approaches, achieving 0.776 mAUC on the HPatches dataset and 0.442 mAUC on the ISC‐HE dataset.