Performance Comparison Between ResNet50 and MobileNetV2 for Indonesian Sign Language Classification

Hearing impairment was considered a significant barrier to understanding verbal communication. Therefore, an alternative communication medium in the form of sign language was required to bridge interactions between Deaf and hearing individuals. One of the sign languages used in Indonesia was the Indonesian Sign Language (BISINDO). The advancement of deep learning technology provided a great opportunity to develop an effective and accurate BISINDO alphabet classification system. This research was conducted to evaluate and compare the performance of two Convolutional Neural Network (CNN) architectures, namely ResNet50 and MobileNetV2, in classifying BISINDO alphabet images consisting of 26 classes from A to Z. Model training wa carried out over 100 epochs and was analyzed using metrics such as training and validation accuracy, precision, recall, F1-score, and confusion matrix. The training process used a dataset that was divided into 80% training data and 20% validation data, and include image preprocessing steps such as resizing and rescaling. The evaluation results showed that ResNet50 achieved 86.42% training accuracy and 98.64% validation accuracy with 98.80% precision, 98.69% recall, 98.57% F1-score, and 31 misclassifications. In contrast, MobileNetV2 showed superior performance with 99.99% training accuracy, 99.65% validation accuracy, 99.69% precision, 99.65% recall, 99.61% F1-score, and only 8 misclassifications. Based on these results, MobileNetV2 was recommended as a more effective and efficient architecture for BISINDO alphabet image classification compared to ResNet50.