Vision-based Static Hand Gesture Recognition using Dense-block Features and SVM classifier
Jaya Prakash Sahoo, Samit Ari, Sarat Kumar Patra · 2022
Recognition of static hand gesture images using a vision-based technique remains a challenging task due to the variations in illumination, shape of the user’s hand, complex backgrounds, and so on. Further, the current convolutional neural network (CNN) technique requires large gesture images to train a model from scratch. Therefore, in this work novel features are obtained from the dense blocks of the fine-tuned DenseNet 201 to recognize the hand gesture images accurately. The proposed feature (denoted as DenseFeat) is a combination of features from last two dense blocks of the fine-tuned DenseNet which gathers enriched information for a gesture image. In addition, to remove the redundant information in the feature vector, the dimension of the DenseFeat is reduced using the principal component analysis (PCA) approach. A support vector machine (SVM) classifier based on a linear kernel is used to recognize the test gesture images. The efficacy of the proposed method is evaluated on two benchmark datasets using both subject-independent and subject-dependent cross-validation techniques. Furthermore, the qualitative and compressive quantitative analysis of the benchmark datasets illustrates that the proposed method outperforms state-of-the-art techniques in terms of mean accuracy and computational time.