Handwritten Language Detection for Low-Resource Languages Using a CNN-BiLSTM Hybrid Model

The proposed study uses a hybrid model that blends convolutional neural networks (CNNs) with bidirectional long short-term memory (BiLSTM). It builds the low-resource language identification system by using a 12,000-image dataset from the real-time dataset, which contains images in various language styles including Tamil, Kurdish, Swahili, and Amharic. To prepare for classification, 3000 images have been gathered for each class. The suggested study uses a BiLSTM for classification and a CNN for feature extraction. Layer pooling reduces the dimensionality of classification by using completely connected layers. The current work used the preprocessed dataset to train CNN hierarchical features across many styles and languages. The accuracy of language detection based on identification reached 97.15%. The main purpose of this model is to recognize graphic writing. The main purpose of this model is to recognize the written language, which helps to preserve historical documents, manuscripts, and manuscripts. In this way, language and cultural heritage are transferred to the future.