Generating Image Captions in Hindi Based on Encoder-Decoder Based Deep Learning Techniques

Image Captioning has experienced significant advancements recently, combining computer vision and natural language processing to create a new field that describes images in words. These approaches utilize an encoder-decoder architecture, where an image is encoded into features by an encoder and those features are decoded into a text sequence by a decoder. Typically, Convolutional Neural Networks (CNNs) are employed as encoders, while Recurrent Neural Networks (RNNs) serve as decoders in these models. Although much of the work in this domain focuses on English, research on Image Captioning models for regional languages is limited. Hindi, being a morphologically rich language and the third most spoken language worldwide, is the focus of this paper. The study conducts a comparative analysis of four state-of-the-art Image Captioning models (ResNet50, InceptionV3, VGG16, and VGG19) specifically applied to the Hindi language. The evaluation of these models’ performance in generating image captions on the widely used Flickr8k dataset employs BLEU, METEOR, and RIBES scores. The results indicate that the InceptionV3 model surpasses the other three models in terms of both BLEU and METEOR scores, making it a valuable reference for researchers operating within this field.