AGR-FCN: Adversarial Generated Region based on Fully Convolutional Networks for Single- and Multiple-Instance Object Detection

Abstract

Addressing the problem that object instance detection has poor detection effect on occluded objects in unstructured environment when using deep learning network, we explore the use of the strategy of adversarial learning in this paper. A three-step pipeline is carried to build a novel learning framework denoted as Adversarial Generated Region-based Fully Convolutional Networks (AGR-FCN). Our method first training the noted deep model Region-based Fully Convolutional Networks (R-FCN), and then an Adversarial Mask Dropout Network (AMDN), which can generate occlusion features for training samples, is designed based on the trained R-FCN. Through the training strategy of adversarial learning between network R-FCN and network AMDN, the ability of network R-FCN to learn the features of occluded objects as well as its instance-level object detection performance is improved. Numerical experiments are conducted for instance detection to compare our proposed AGR-FCN with the original R-FCN on the self-made BHGI Database and the public database GMU Kitchen Dataset, which demonstrate that our proposed AGR-FCN outperforms original R-FCN and can achieve an average detection accuracy of nearly 90%.