Özet:

The main transmission routes of COVID-19, an international public health emergency, are respiratory droplets and physical contact. As one of the comprehensive strategies to prevent and fight against disease the outbreak, the use of medical masks in the public sphere has been made mandatory in many societies. In this context, automatic control of the use of medical masks in the public sphere is crucial in the fight against the outbreak. This study aimed to detect the use of medical masks automatically from images of the public sphere by the transfer learning approach. By transfer learning approach to deep architecture, it is aimed to obtain effective solutions in medical mask detection with fine-tuning of pre-trained parameters. The image data set offered by Human in the Loop (HITL) as open access was used for the automatic detection of medical masks. The SqueezeNet based on transfer learning approach proposed in this study achieved a classification accuracy of 99.20%. In addition, the AUC (area under the ROC curve) value was found as 0.998. To emphasize the superiority of the transfer learning approach, the SqueezeNet architecture, which does not contain trained parameters, was also applied to the same data set and the obtained performance metrics were compared. The model trained from scratch on a limited number of image dataset offered classification accuracy and AUC performances as 94.75% and 0.976, respectively. As a result, it has been observed that the deep architecture, which has been trained in a very short time with the transfer learning approach, has an impressive performance in detecting the use of medical masks.

Anahtar Kelimeler: