Face images in digital archives have a large age span and suffer from varying degrees of image degradation over time, leading to a significant degradation in the performance of generic face recognition models. To address the above problems, this paper proposed an anti-noise cross-age face recognition model. The model combines local residual learning and soft thresholding module, then embeds them into the backbone network to remove irrelevant features and guide the network to extract valid initial facial features. In this case, the soft thresholding module adaptively sets thresholds by branching at two different scales. The initial facial features are decomposed into age-related features and identity-related features. The identity-related features are used for face recognition. Meanwhile, a benchmark test dataset based on real archives was constructed in this paper. The study shows that the model has a high degree of robustness and anti-noise interference. Also, soft thresholding has a positive impact on noise suppression.