PCA-SVM for a Lightweight ASL Hand Gesture Image Recognition

Other studies have previously researched creating a new dataset called Gesture Modified National Institute of Standards and Technology (MNIST Gesture), similar to the MNIST dataset for American sign language (ASL) hand gestures. Several methods, such as long short-term memory (LSTM), have been used for gesture detection on the dataset. However, there is a research opportunity to provide a learning model that requires less training data. Our research aim is to use PCA-SVM for lightweight hand gesture recognition. The first step of our research is to collect the MNIST Gesture dataset from Kaggle. Then, we do preprocessing and prediction model training. We then compare PCA with the mean decrease in impurity (MDI) and the least absolute shrinkage and selection operator (LASSO) method in their performance to reduce dataset dimensions in the SVM training process. Lastly, we compare our proposed model with two state-of-the-art benchmark models: long short-term memory (LSTM) and convolutional neural network (CNN). The test results show that the optimum number of PCA-SVM principal components in ASL hand gesture recognition is 50. Then, PCA performs better than MDI and LASSO on dimension reduction with an SVM model. Finally, the PCA-SVM model bests LSTM in hand gesture recognition by accuracy, precision, recall, and f1-score. However, PCA-SVM and CNN both have an accuracy higher than 0.99. On the other hand, PCA-SVM is more lightweight than CNN because it has a shorter training time and testing time, namely 3.5 s and 0.2 s, respectively.