Comparison of Transfer Learning Algorithm Performance in Hand Sign Language Digits Image Classification

Abstract

Image hand sign classification has become an interesting topic in image processing and machine learning. However, to achieve optimal performance in hand sign image classification tasks, a large and diverse dataset as well as powerful learning algorithms are required. One popular technique for improving the performance of classification models is transfer learning, which allows the use of knowledge learned from previous models and applies it to new tasks. In this study, the performance of two different transfer learning algorithms, ResNet-50 and VGG-16, was compared on the Sign Language Digits Dataset, which consists of 10 different types of handwriting images. The results of the experiment showed that both tested transfer learning algorithms had good performance. However, VGG-16 provided the best results with an accuracy of 97,29%, precision of 97,38%, recall of 97,45%, and an F1 score of 97,36%, while ResNet-50 achieved an accuracy of 94,57%, precision of 94,75%, recall of 94,96%, and an F1 score of 94,78%. In conclusion, transfer learning algorithms are effective techniques for improving the performance of hand sign image classification models. Choosing the appropriate transfer learning algorithm and dataset can help generate more accurate classification models.