Captioning Remote Sensing Images Using Transformer Architecture

Image Captioning aspires to achieve a description of images with machines as a combination of Computer Vision (CV) and Natural Language Processing (NLP) fields. The current state of the art for image captioning use the Attention-based Encoder-Decoder model. The Attention-based model uses an 'Attention mechanism' that focuses on a particular section of the image to generate its corresponding caption word. The NLP side of this model uses Long Short-Term Memory (LSTM) for word generation. Attention-based models did not emphasize the relative arrangement of words in a caption thereby, ignoring the context of the sentence. Inspired by the versatility of Transformers in NLP, this work tries to utilise its architecture features for the Image Captioning use case. This work also makes use of a pretrained Bidirectional Encoder Representation of Transformer (BERT) which generates a contextually rich embedding of a caption. The Multi-Head Attention of the Transformer establishes a strong correlation between the image and contextually aware caption. This experiment is performed on the Remote Sensing Image Captioning Dataset. The results of the model are evaluated using NLP evaluation metrics such as Bilingual Evaluation Understudy 1-4 (BLEU), Metric for Evaluation of Translation with Explicit ORdering (METEOR) and Recall-Oriented Understudy for Gisting Evaluation (ROUGE). The proposed model shows better results for a few of the metrics.