Deep Reinforcement Learning For Object Detection

1Satya ranjan Das, Sarita Mohapatra

178 Views
33 Downloads
Abstract:

This paper present a technique for performing object detection in images guided by deep reinforcement learning. The key idea is to specialize in those parts of the image that contain richer data and zoom on them. This paper tend to train an intelligent agent that, given an image window, is capable of deciding wherever to focus attention among five completely different predefined region candidates (smaller windows). Distractors usually obstruct the object of interest and cause it to disappear from the field of view. This paper tend to propose hand/eye controllers that learn to move the camera to stay the object inside the field of view and visible, in coordination to manipulating it to realize the specified goal, e.g. pushing it to a target location. This paper tend to incorporate structural biases of object-centric attention inside our actor-critic architectures, that our experiments recommend being a key permanently performance. Our results additional highlight the importance of the information with reference to the environmental problems. During this paper, this paper tends to present principled successive models that accumulate proof collected at a little set of image locations to detect visual objects effectively. By formulating consecutive search as reinforcement learning of the search policy (including the stopping condition), our totally trainable model will explicitly balance for every category, specifically, conflicting goals of exploration sampling a lot of image regions for higher accuracy, and exploitation stopping the search with efficiency when sufficiently confident concerning the target’s location.

Keywords:

Image Detection, Computer Vision, Reinforcement Learning, Deep Learning

Paper Details
Month12
Year2019
Volume23
IssueIssue 5
Pages255-261