A Deep Reinforcement Learning Framework for Multi-Stage Optimized Object Detection

Nowadays, with the significant development of artificial intelligence and especially after the emergence of deep learning, vision systems specifically the vision of robots have become more interesting. One of the most important challenges of computer vision and robotics is object detection. In fact, high computational complexity bothers object detection techniques that rely on region proposals or predefined anchors. We propose an optimized method upon a deep reinforcement learning framework along with multi-stage approaches for object detection as well as train two intelligent agents for a more accurate search of objects and localize them in images. Several candidate bounding-boxes are proposed to the algorithm in the first stage, and the perfect boxes are selected for the final boxes around the objects in the second stage using an optimal approach. Our approach was assessed using two renowned datasets in the object detection field, PASCAL VOC 2007 and PASCAL VOC 2012. According to these evaluations and experiments, the proposed method not only outperforms similar non-region proposal methods but also does well detecting small objects.