The problem of combining embeddings from set of face images for person identification is considered. The compact representation of multiple embeddings reduces the computational complexity of comparing person identifiers, and takes into account the variability of face images taken during registration and identification. To improve the quality of presentation, it is proposed to use the weight aggregation of embeddings with weights determined by the trainable attention module. The proposed DoubleQualityGhostVLAD (DQG_VLAD) attention module is implemented by a combination of GhostVLAD method with local clustering [1] and Multicolumn Network model [2]. The DQG_VLAD approach evaluates the weights of image embeddings for aggregation by constructing their clusters with an assessment of their relative quality and importance in terms of the classification loss function. The presented results show that the proposed method increases the accuracy of facial images identification from 1% to 4% compared with the basic algorithms.