Konumsal Bilgi Evrisimsel Sinir Aglarini Nasil Etkiler?

The success of Transformers, including in the field of image processing, has recently attracted the attention of researchers. Some of the researchers tried to design Transformer- based or Convolutional Neural Network-based models separately, while others tried to combine them to produce hybrid models. A significant amount of hybrid model studies have examined the sub-model design and/or the applicability of self-attention in Convolutional Neural Networks. However, position embedding, another contribution from Transformers, has received much less attention. In this study, the effect of position information on Convolutional Neural Networks was analyzed. As a result of the experiments, it has been observed that the use of position information affects performance. In the AgeDB-30, CALFW, and LFW test sets, models with different position information usage have been able to surpass the performance of the model without position information by achieving 95.12%, 93.95%, and 99.52% accuracy, respectively.