融合对抗学习策略的实例级物体检测算法  

Instance-Level Object Detection Algorithm Fusing Adversarial Learning Strategies

作　　者：覃润楠 王睿[1] QIN Runnan;WANG Rui(School of Instrumentation and Optoelectronic Engineering,Beihang University,Beijing 100191)

机构地区：[1]北京航空航天大学仪器科学与光电工程学院

出　　处：《模式识别与人工智能》2019年第11期1006-1013,共8页Pattern Recognition and Artificial Intelligence

基　　金：国家自然科学基金项目(No.61673039)资助~~

摘　　要：针对目前基于深度学习的实例级物体检测算法对受遮挡物体的检测效果较差的问题,文中引入对抗学习的训练策略,提出改进的对抗生成式区域全卷积网络算法(AGR-FCN).以区域全卷积网络(R-FCN)为基准框架,添加为训练样本生成遮挡特征的对抗性遮挡丢弃网络(AMDN).通过R-FCN与AMDN间对抗学习的训练策略,提升R-FCN对遮挡物体的特征学习能力,优化整体实例级物体检测性能.在公共数据库GMU Kitchen和自制数据库BHGI上的实验表明,在复杂多变的非结构化环境中,如随机变化的不同光照、尺度、焦比、视角与姿态、遮挡等条件下,AGR-FCN的平均检测精度较高.Existing instance-level object detection algorithms based on deep learning achieve a poor detection effect on occluded objects.To solve the problem,an improved adversarial generated region-based fully convolutional networks(AGR-FCN)with the training strategy of adversarial learning is proposed.The original fully convolutional networks(R-FCN)is regarded as a fiducial frame,and adversarial mask dropout network(AMDN)is designed based on the trained R-FCN to generate occlusion features for training samples.Through the training strategy of adversarial learning between R-FCN and AMDN,the learning ability of R-FCN to the features of occluded objects is improved,and its overall instance-level object detection performance is optimized.Experiments on GMU Kitchen dataset and BHGI dataset show that AGR-FCN algorithm achieves good detection accuracy in complex and changeable unstructured environments,such as randomly varying illumination,scale,focal ratio,angle and attitude and occlusion.

关 键 词：实例级物体检测 对抗学习 区域全卷积网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]