Cryo-electron microscopy (cryo-EM) is a pivotal technique for elucidating protein structures, yet particle picking remains a bottleneck due to inherent challenges such as specimen impurities, sample preparation variability, and microscope parameter fluctuations. These factors result in micrographs with diverse noise profiles, pixel characteristics, and particle dimensions, posing significant hurdles for conventional supervised methods that struggle with generalization and necessitate labor-intensive expert annotations. In this work we present a self-supervised method that leverages a Masked AutoEncoder's representation space to sequentially denoise micrographs based on clusters with different noise levels. Evaluation across 14 datasets demonstrates superior generalization capabilities compared to state-of-the-art supervised methods, showcasing consistent performance independent of pre-training data. This underscores self-supervised learning's potential for advancing cryo-EM image analysis and enabling more efficient structural biology research.