Imageem: Pre-Trained Encoded Vector Embeddings for Image Modelling

Abstract

Today some of the most intriguing challenges posed in the field of AI, is the lack of computational resources required to train the deep learning models, especially the image related problems. These Image modelling problems are solved by training a Convolutional Neural Network (CNN) which is computationally expensive process. Our aim is to extract the features of the images by training on these vector embeddings locally and then deploying it in the server for easy access to the researchers. The proposed system provides high dimensional vectors that capture dense features of objects. Pre-trained word embedding is a common way of representing words in a vocabulary present in a document. These embeddings have the capability to capture the context of a particular word in a sentence or an entire document, in relation with words other than the one under review. A popular algorithm for training these word embeddings is skip-gram model proposed in Word2Vec architecture.By using these trained vector embeddings of popular image models, researchers can download them and apply it to their use case which will reduce their training time by a thousand fold. Our aim is to make it so simple for the user that he can train it on edge devices like raspberry pi, jetson nano, these pre-trained vectors image models of high accuracy can be trained on even mobile phone processor.