Published Paper Details:

The latest developments in Deep Learning based Machine Translation and Computer Vision based Object Detection have led to high accuracy Image Captioning models. Although these models are very accurate, they tend to rely on the use of expensive computational power making it difficult to use these models in real-time applications like processing the video streams in real time and extracting the information. In this paper, we carefully follow some of the heuristic strategies and core ideas of Image Captioning and its common methods and present our simple sequence to a sequence based implementation with a remarkable transformation and efficiency such as using beam search instead of greedy search that allows us to implement these on low-end hardware. The proposed system compares the results calculated using a variety of metrics with high-quality models and analyses the reasons behind the model trained on the MS-COCO dataset that are lacking due to trade-off between computation speed and quality. In this proposed system, Restful API endpoint will be created to be used on any device with an internet connection such as a mobile phone, IoT devices, clock, etc, this endpoint used to sent an image to the model running on remote server which in response will generate and sent an caption describing the objects and their relationship with each other in image in a natural language.