PROJECT TITLE :

CNN Fixations An Unraveling Approach to Visualize the Discriminative Image Regions

ABSTRACT:

Vision research has been transformed by deep convolutional neural networks (CNN), which have been adopted for a wide range of applications, including classification and detection. The inner workings of these devices are generally kept a secret because of the lack of transparency they provide. In this project, we intend to provide visual explanations for the network's predictions in order to alleviate this opaqueness of CNNs. CNN-based models trained for applications such as object recognition and caption generation can be analysed using our method. We do this by unravelling the forward pass process, which is a novel approach. Discriminative picture locations can be discovered by exploiting feature relationships across layer hierarchies, which can then be used to train an artificial neural network (ANN). They are called CNN-Fixations, which is a loose analogy for human eye fixations. There are no architectural changes, additional training or gradient computations required to implement our solution (CNN Fixations). Different network designs, multiple vision challenges, and different input modalities can be used to demonstrate that our approach is able to locate the discriminative picture spots.