Işitme Engelli Çocuklar için Özelleştirilmiş Hafif Ağırlıklı Yüz Ifadesi Tanıma Modeli

This study aims to use the advantages of the lightweight model in recognizing facial expressions of hearing-impaired children. For this purpose, the weight pruning technique is used and the values of unnecessary parameters were zeroed out. Also, pruning is applied during training to prevent the model from experiencing a decrease in facial expression recognition performance. The results showed that there is no decrease in facial expression recognition performance when pruning is applied with a VGG-Face based model on the CAFE and HIC (hearing impaired dataset) datasets. In addition, efficiency has been gained in the storage area of the models. The model evaluated on the HIC dataset proved its efficiency regarding inference speed with an average prediction time of 5.67 ms (standard deviation: 0.44 ms) per sample in TensorRT format. With this study, for the first time in the literature; a lightweight deep learning model trained with the weight pruning method is used to recognize facial expressions of hearing-impaired children.