Dynamic Routing Using Inter Capsule Routing Protocol between Capsules

A capsule is a combination of multiple neurons, which designed to analyze specific feature representations in an image, a capsule network resembles with CNN multiple layer network model, in which convolutional and ReLu function is followed by max pooling which help in reducing the input values for processing. However, we observed that for each capsule in layer l to layer (l +1), we need to calculate a prediction vector, whose magnitude defines the detected feature in an image and the orientation define the pose of the object in an image. Now calculating the prediction vector for each capsule in the lower layer is prolonged. This is because, in capsule network, we are not securing the gained knowledge for future use. In this paper, we are introducing a protocol named as ICRP (Inter Capsule Routing Protocol), which uses FDM (Feature Detection Matrix) to store the learned information at every iteration in the capsule network. This reduces the processing time for feature detection in an image gradually, by searching the similar feature in FDM before processing the network.