Recently, metric learning has been used in various applications. In this type of learning, the reduction of dimension is done to the quality that similar samples are closer to each other in the final subspace than different samples; that is, it defines a similarity criterion for samples based on the distances in the final subspace. Due to the special function of metric learning, it has been used in many applications in the past years. In this research, a metric learning method based on deep neural networks has been designed to define a similarity criterion. Since the similarity criteria is different when faced with multi-modal datasets, a deep clustering method has been used along with deep localized metric learning. Bearing this in mind, a new architecture has been designed based on deep neural networks. It is worth mentioning that in this study, the relationship between deep clustering and deep meter learning is considered end-to-end. To eva‎luate the proposed method, Viper datasets and a combination of Viper and QMUL datasets have been used. The samples in both datasets are full-length images front and back of humans in the subway and park environment. The small number of examples and the low quality of images 48 by 128 are the main properties of datasets that have created challenging conditions. Finally, according to the rich accuracy of this research compared to previous studies, the final similarity criterion more specifically covers nonlinear relationships between samples.

مهران صفاياني

عبدالرضا ميرزائي

الهام محمود زاده، مازيار پالهنگ