Abstract:
Following the breakout of COVID-19, there has been a significant increase in demand for
gesture sensing applications, which allow users to manage gadgets with simple hand gestures
rather than physically touching them. In comparison to expressions, actions, and other
interaction techniques, gestures are more intuitive, straightforward, and natural. For Smart
AI assistants, a cognitive vision system is essential for enabling seamless interaction with
humans. Among the various features in the vision system, the ability to detect and recognize
different hand gestures provides significant value addition. Systems employing hand gesture
recognition technology are capable of distinguishing specific gestures such as victory sign,
thumbs up, wave, peace sign, rock sign, number counting, etc. Real-world systems designed
for human-computer interaction struggle to recognise and categorize hand gestures because
1) People perform gestures in a wide variety of ways, based on their cultural difference, 2)
Variability of input lighting, distance limits, 3) Requirements of larger datasets etc. The
wearable glove-based sensor technique and the camera vision-based sensor approach are the
two main strategies for hand gesture recognition research. In this work, images captured
using camera sensors are used as the input and fed into the proposed shallow convolutional
neural network for classification and prediction of hand gestures. The proposed low weight
convolutional neural network achieved faster training results by using fewer parameters and
training epochs. Three well-known pre-trained models, including VGG16, ResNet50, and
Mobilenet, are also taken into consideration for comparison. These models are applied
on Fingers number count and LeapGestRecogn dataset and evaluation measures such as
F1 score, recall, accuracy, and precision are computed to analyze the performance of the
models.The experimental results indicate that the proposed model has achieved a recognition
rate of 99.9% and 99.79% on LeapGestRecogn and Fingers number count dataset respectively.
Furthermore, the proposed model outperformed the state-of-the-art methods with better
accuracy
