Optimized Emergency Sign Language Recognition with Large-Kernel Attention Graph Convolutional Network for Accurate Gesture Classification

Sign language recognition (SLR) plays a significant role in improving communication accessibility for hearing-impaired individuals, particularly in emergency scenarios where fast and reliable gesture interpretation is required. However, emergency Indian Sign Language (ISL) recognition remains challenging due to rapid hand movements, visually similar gestures, and complex spatial-temporal relationships among hand components. This study proposes an SLR-HIE-LKAGCN framework that integrates topology-aware spatial representation, spatial-temporal feature learning, graph-based gesture modeling, and optimization-based parameter adaptation for emergency ISL recognition. Initially, input gesture frames are processed using a Simplicial Convolutional Filter (SCF) to preserve higher-order spatial relationships among hand regions and enhance gesture representation. The refined gesture sequences are then provided to a Spatial-Temporal Knowledge-Embedded Transformer (STKET) to capture temporal dependencies and contextual movement patterns. The extracted features are further analyzed utilizing a Large-Kernel Attention Graph Convolutional Network (LKAGCN), which models relationships among hand landmarks and learns discriminative gesture representations. In addition, the Secretary Bird Optimization Algorithm (SBOA) is employed to optimize critical network parameters and improve model convergence. The proposed framework is evaluated on emergency ISL gesture categories, including accident, call, doctor, help, hot, lose, pain, and thief. Experimental outcomes establish that the proposed technique achieves superior recognition performance with 98.82% accuracy, 98.57% precision, and 98.53% F1-score, outperforming existing approaches. The obtained results indicate that SLR-HIE-LKAGCN provides an effective and computationally efficient solution for real-time emergency ISL recognition.