Prediction of three-dimensional structures of proteins from amino-acid sequences is an important problem of bioinformatics. We still do not have, however, a reliable computational method for de novo prediction, i.e., prediction of structures that do not have homologues of known structures. One way to improve techniques of de novo prediction is to assess quality of the predicted model structures to select candidate models from them before knowing the experimentally determined structure. In this paper, we develop a new method of model quality assessment for de novo prediction, the fragment-based consistency score (FCS) method. In the FCS method, fragments from library proteins are collected for each fragment of the model, and structural similarities between the model fragment and the library fragments are measured for assessing qualities of the model. Three structural indices are employed; secondary structure, local density and local contact order. The optimal performance was obtained when relatively correlated fragments are collected, and structural similarity is measured in fuzzy comparison and averaged by finite width matching. The FCS method can select partially correct models for hard targets of de novo prediction examined in CASP7 and CASP8. We also evaluated the abilities of the FCS method to select structural models of loops or coil regions in targets, and to distinguish proteins which are similar in sequence but have greatly different conformations from each other. The proposed FCS method helps to improve the de novo prediction scheme and is also useful to solve difficult structures in the template-based modeling.