EfficientNetB0-CBAM: An attention-guided framework for multilingual script identification

Abstract

Script identification remains challenging for several applications, including handwritten text recognition, scene text analysis, and low-resource languages such as Kashmiri, for which benchmark datasets for standardization are limited. The study addresses this challenge by collecting a multilingual dataset of Kashmiri, Urdu, and English script images and proposes a deep learning model that uses the Convolutional Block Attention Module (CBAM) to learn the hidden and overlapped script features. The proposed EfficientNetB0-CBAM model was evaluated against the baseline EfficientNetB0 model, ResNet50, MobileNetV2, and EfficientNetB0 with a Squeeze-and-Excitation (SE) module. Experimental results showed that the proposed EfficientNetB0-CBAM framework outperformed all compared models, achieving an accuracy of 93.32 ± 0.71% compared to 91.28 ± 0.85% for the baseline EfficientNetB0 model, while also outperforming ResNet50, MobileNetV2, and EfficientNetB0 with the SE module. Moreover, the framework was comprehensively evaluated using several performance metrics, including Balanced Accuracy, ROC-AUC, training-validation accuracy-loss curves, Macro-F1, and confusion matrix analysis. These results show that incorporating the attention mechanism enhances feature representation and classification performance, demonstrating its effectiveness for multilingual script identification in low-resource scenarios.