SqueezeNextV2: Rethinking SqueezeNext Deep AI model to Detect American Sign Language (ASL)

Aamosh B. Thapa
Prof. Nazmul Shahadat, Faculty Mentor

This paper presents SqueezeNextV2, an improved lightweight deep learning model for real-time American Sign Language (ASL) hand gesture recognition. ASL is a vital communication tool for the deaf and hard-of-hearing community, yet existing recognition systems are often too computationally intensive for mobile and embedded devices. To address this, we enhance the SqueezeNext architecture with targeted modifications, including a corrected 29-class output layer, dropout regularization, label smoothing, adaptive average pooling, stronger data augmentation, cosine annealing learning rate scheduling, and Kaiming initialization. The model is trained on a diverse dataset of 18,468 images and evaluated on a held-out test set. SqueezeNextV2 achieves a test accuracy of 97.15% with only 585,341 parameters, demonstrating strong generalization without overfitting. These results highlight the effectiveness of combining efficient architectures with principled training strategies for accurate, real-time ASL recognition on resource-constrained devices.

Keywords: American Sign Language (ASL), Deep Learning, Convolutional Neural Networks (CNN), SqueezeNext, Gesture Recognition, Computer Vision, Lightweight Models, Edge Computing

Topic(s):Computer Science
Linguistics

Presentation Type: Oral Presentation

Session: -
Location: MG 2007
Time: 

Add to Custom Schedule