Using single particle images taken by electron microscopy, pH-dependent conformational changes of human serum albumin were investigated. Despite the noisy particle images of negatively stained serum albumin (67 kDa), our novel algorithm for automated particle picking and reference-free classification resulted in the appropriate grouping of the particle images. Iteratively aligned particle images in the same group provided recognizable image features for individual groups. In a pH 7.0 study, monomer images were consistent with an available crystal structure model; the dimer images were separated into different classes. At pH 3.5, the monomer images were similar to those at pH 7.0; slight differences included a small number of elongated conformations and increased population of larger multimers. Our images were also compared with projection images of an atomic model from crystallography, and demonstrated consistency of the molecular conformation both at pH 7.0 and pH 3.5. Our classification method was effective in discriminating monomers from a mixture of different conformations of the protein, enabling the study of the conformational dynamics of small proteins, using the atomic model as a reference.