Deep Neural Network-Based Sign Language Translation Model with Dynamic Coordinate Compression

• Min-Jae Jeong,
• Soonghwan Ro,
• Jun-Ki Hong,

Vol. 19, No. 7, July 31, 2025
10.3837/tiis.2025.07.003, Download Paper (Free):

• deep neural network
• Dynamic Coordinate Compression
• Sign Language Translation

Abstract

This study proposes a deep neural network (DNN)-based Dynamic Sign Language Coordinate Compression (DDSC) model that significantly reduces training time while preserving the semantic integrity of sign language during training and translation. The DDSC model analyzes the Euclidean distances between coordinates extracted from each video frame and selectively compresses segments that exhibit repetitive or minimal motion, while preserving frames with substantial movement that are crucial for conveying meaning. Frame intervals with low inter-frame distance are interpreted as repetitions of the same sign and are compressed to reduce redundancy. In contrast, frames with larger movement variation are retained to prevent semantic loss. The proposed DDSC model was evaluated using a Korean Sign Language (KSL) dataset consisting of word-level video samples for 16 frequently used signs, collected from the AI-Hub platform. The experiment compared three types of coordinate input data: uncompressed, uniformly compressed, and dynamically compressed using DDSC. The corresponding translation accuracies were 0.945, 0.776, and 0.853, respectively. Compared to uniform compression, the proposed DDSC model achieved approximately a 60.73% reduction in training time with only a 9.74% drop in accuracy. These results demonstrate that the DDSC model offers a practical trade-off between accuracy and computational efficiency, making it well-suited for deployment in real-time or resource-constrained environments such as mobile or embedded systems.

Statistics

Cite this article