Noise-Robust Face Recognition via Non-target Similarity Distribution Guided Sample Selection

Label noise is a major challenge in large-scale supervised face recognition, where weak or automatic annotations often introduce errors that mislead training. To address this, we propose a noise robust framework that performs noise detection and sample selection directly in cosine-similarity space. We observe that the non-target cosine similarities of clean samples share a highly consistent distribution profile with the target similarities of unfitted mislabeled samples. This phenomenon can be explained from a backpropagation perspective and provides a cue for monitoring label noise during training. Based on this observation, we formulate noise detection as boundary estimation in similarity space. Specifically, we track the upper bound of high-confidence clean nontarget similarities to determine the filtering threshold, without requiring prior knowledge of the noise rate or auxiliary networks. We further introduce a progressive rule during early training, where the threshold gradually increases from the estimated noise lower bound to the upper bound. This process discards unreliable samples while retaining hard but clean samples. Extensive experiments on eight synthetic and three real-world noisy datasets demonstrate that our method achieves superior noise detection and state-of-the-art recognition accuracy, with only about 0.3% average measured overhead. The filtered dataset produced by our method is also beneficial for subsequent training. Source code is available at https://github.com/wfl95/DDLN.