Q-DFN

Abstract

Listen

Translate article

Traditional approaches to each image simply depend on detectors of the classification network and the bounding box regression network. These approaches are, although working well, fail to exploit the satisfied accuracy for specific object detection or many methods sacrifice the accuracy for the speed of real-time detection. The reason for the unsatisfied results is that the relationship between the various objects is not effectively considered. In this paper, we propose a decisive fusion network for object detection. After the network extracts the features from the backbone first, the last five layers are treated as the feature pyramid structure. Each layer is responsible for the bounding box by the border regression algorithm, and the Deep Q-Learning (DQN) algorithm is responsible for the classification of the target. The fusion function connects the DQN algorithm with the object detection network, and considers the selection relationship between objects through the DQN algorithm to improve the overall accuracy. Through a lot of experiments, the parameters of DQN are better adjusted. The experiments show our method is better than tradition approaches.