Research Article
Developing an Efficient and Lightweight Deep Learning Model for an American Sign Language Alphabet Recognition Applying Depth Wise Convolutions and Feature Refinement
@INPROCEEDINGS{10.4108/eai.28-4-2025.2357809, author={Pillarisetty Uday Karthik and Sai Subbarao Vurakaranam and Sumalatha M and Renugadevi R and Sunkara Anitha}, title={Developing an Efficient and Lightweight Deep Learning Model for an American Sign Language Alphabet Recognition Applying Depth Wise Convolutions and Feature Refinement}, proceedings={Proceedings of the 4th International Conference on Information Technology, Civil Innovation, Science, and Management, ICITSM 2025, 28-29 April 2025, Tiruchengode, Tamil Nadu, India, Part I}, publisher={EAI}, proceedings_a={ICITSM PART I}, year={2025}, month={10}, keywords={artificial intelligence convolutional neural networks deep learning dental caries detection oral imagery}, doi={10.4108/eai.28-4-2025.2357809} }- Pillarisetty Uday Karthik
Sai Subbarao Vurakaranam
Sumalatha M
Renugadevi R
Sunkara Anitha
Year: 2025
Developing an Efficient and Lightweight Deep Learning Model for an American Sign Language Alphabet Recognition Applying Depth Wise Convolutions and Feature Refinement
ICITSM PART I
EAI
DOI: 10.4108/eai.28-4-2025.2357809
Abstract
In this, we proposed a deep learning framework for classifying American Sign Language (ASL) alphabet gestures to support accessibility for people with speech and hearing impairment. We evaluated our model using three public ASL datasets. one of the datasets of 87,000+ real-time images and 29 classes. To establish a baseline, we also tried state-of-the-art models, including VGG16, Efficient Net, MobileNetV1/V2, and ResNet50. Modeled after these findings, we created a compact application-specific convolutional neural network (CNN) model for static ASL recognition. Our custom ASL alphabet model employs depthwise separable convolutions, batch normalization, dropout, and global average pooling and adds a loop-based feature refinement block that is executed four times to increase spatial feature learning. Our test runs showed our model consistently achieved over 95% across various datasets while being faster and more reliable than the bulkier models. This model is a great candidate for real-time applications.
