The paper presents a description and justification of the correctness of fuzzy recognition by a logic-predicate network. Such a network is designed to recognize complex structured objects that can be described by predicate formulas. The NP-hardness of such an object recognition requires to separate the learning process, leaving it exponentially hard, and the recognition process itself. The learning process consists in extraction of groups of features (properties of elements of an object and the relations between these elements) that are common for objects of the same class. The main result of a paper is a reconstruction of a logic-predicate recognition cell. Such a reconstruction allows to recognize objects with descriptions not isomorphic to that from a training set and to calculate a degree of coincidence between the recognized object features and the features inherent to objects from the extracted group.