Efficient Object Detection using Deep Reinforcement Learning and Capsule Networks

Recently, with the advent of deep learning, significant advances in computer vision have been made. One of the traditional and important tasks in computer vision is object detection. Object detection methods based on pre-defined anchors or region proposal suffer from high computational complexity. We propose a new method based on deep reinforcement learning and capsule networks for object detection in images. Capsules are a group of neurons that communicate with each other in the form of routing. The main idea is to use capsule networks as the heart of deep reinforcement learning and train an intelligent agent for a more accurate search of objects and localize them in the image. We also defined a new function called CEU and used it as part of the movement reward function. We evaluated our method on two well-known object detection benchmark datasets called PASCAL Visual Object Classes (VOC) 2007, and 2012. Experiments illustrate that the proposed method achieved higher precision than similar methods in terms of not being a region proposal.