In this paper, we propose a novel framework for gait quality assessment based on image analysis, extending the traditional signal-based approach. Specifically, we construct Cycle Dissimilarity Images (CDI) from raw gait signals. Such images summarize all local dissimilarities existing in the dynamics between a gait signal and one normal gait reference. Also, we construct a typical dissimilarity image, by matching each normal gait reference to itself. Then, we propose to quantify gait deviations by computing the distance between the CDIs and the typical dissimilarity images. Our results indicate that, compared to the signal-based approach, this new framework leads to a more precise gait deviation assessment, and a more refined characterization of motor impairments, as hemiparesis, tetraparesis, and paraparesis.