Creating Shareable Instagram-Like Image Filters
This report presents a POC where we propose a novel way to learn image transformations using neural networks. Such a technique can be used to learn filters such as the ones found in Instagram. With the process that we propose, a user can define their own image filter using a set of transformations (changing color intensity, vignette effects etc.) on a single image. Thereafter, these set of transformations are applied on randomly created images which act as a training set, and these random images are then used to train the network. With this technique, we reduce the overhead of storing a large training set, or asking a user to define their transformations on multiple images. The trained neural network along with its weights can be transferred to a third party application (such as Instagram), and then reused in a seemless manner. The network size can be further reduced by using network compression techniques, which are left as future work.