Context-Based Quote Generation from Images

Abstract

Photographs are compact stores of special moments of our lives. Due to the convenience of photography tools such as mobiles and cameras, we can capture each and every moment of celebration. We often share those moments with our friends and colleagues and sometimes put them on our social media profiles. To make our social media posts attractive, we usually try to associate them with some inspirational quotes. Those quotes are either self-created or searched from the Internet. Searching and selecting the right quote from the Internet for our photograph is tedious as well as uninspiring at the same time. We have to first think about the theme which our image is depicting and then we have to search quotes related to that theme. After this, we have to filter out the most probable quotes for our image from about thousands of quotes. Our application aims to ease the process of searching and associating the right quote for the image. It accepts a picture as an input and after processing it, suggests relevant quotes for the image to the user. To achieve this functionality, we have used the computation power of Convolution neural networks, the concept of Long Short Term Memory and Similarity measures for suggesting the suitable caption for the image which are then further utilized to render quotes to the user. In this paper, we not only represent the method of conversion of image to quote but also a comparative study on performance measures (BLEU scores and validation loss in this work) of various pre-trained CNN models, and based on the comparative study, the best performing CNN model is chosen (MobileNetV2 in our study).