Sign Language Detection System using MobileNetV2: A Deep Learning Approach for Real- Time Recognition

Sign language detection systems will be of great importance in facilitating communication between an individual with hearing impairments. Other unique aspects of the current article are the suggested pipeline of real-time sign language detection namely, MobileNetV2, a lightweight deep learning architecture developed to perform in the mobile and resource-constrained systems. Our system uses the efficiency of the depthwise separable convolutions and inverted residual blocks in MobileNetV2, to obtain a good degree of accuracy, despite being efficiency-wise very much computationally. We have created a multimedia dataset to train (i.e., 24 alphabets of the American Sign Language, or simply ASL) and have used the idea of transfer learning to customize a previously trained MobileNetV2 model to identify the sign language. The proposed system is then tested on our test dataset with accuracy of 94.7 percent and low running costs of running the proposed system. Memory (3.4MB)/ processing (35 FPS) mobile. The output of experiments confirmed the effectiveness of our method as it outperforms standard CNN networks in terms of both accuracy and speed of calculations, i.e., it could be used in real life. The versatility of the system via its real time ability to be used even on mobile devices opens up new horizons of convenient to operate communication modes to the deaf and hard of hearing people.