Abstract:
Script identification is an important task in document analysis and recognition systems, which involves determining the script or writing system of a given text. In many multilingual countries, such as India and Bangladesh, documents often contain text written in different scripts, including Bangla and Devanagari scripts. However, identifying the script of degraded text poses significant challenges due to factors such as noise, distortion, and degradation caused by various environmental conditions and scanning artifacts. Using the VGG16 and CNN (Convolutional Neural Network) architectures, we offer a novel method in this research for identifying degraded scripts in Bangla and Devanagari scripts. Convolutional layers are used in the neural network design known as the VGG16, which has demonstrated success in image recognition. On the other hand, convolutional neural networks (CNNs) are commonly utilized for text recognition tasks because they are skilled at identifying local features. The proposed approach leverages the power of both VGG16 and CNN to effectively identify the script of degraded text. We conduct analyses on a standard dataset made up of erroneous Bangla and Devanagari scripted text samples. In order to increase the robustness of our model, we subjected samples from the dataset to different levels of degradation, such as blurring, noise, and distortion. To improve the model’s performance on the preprocessed dataset,
we employed transfer learning methods by using a pre-trained VGG16 model as a feature extractor for the CNN model. This helped to enhance the performance of both models on the dataset. Our proposed approach yielded better results than existing methods for identifying degraded script in Bangla and Devanagari scripts, according to our experiments. Our approach also achieved state-of-the-art performance. The combined use of VGG16 and CNN significantly improves the accuracy of script identification compared to using only one of the architectures. In addition, the proposed technique has been demonstrated to be more effective than current methods in terms of its ability to handle degradation and maintain robustness. This means that the proposed approach can efficiently manage different types of degradation that are usually present in real-world situations.