基于生成对抗网络学习被遮挡特征的目标检测方法

Object detection via learning occluded features based on generative adversarial networks

实际生活中目标间存在的遮挡会造成待检测目标的特征缺失,进而使得检测准确度降低.鉴于此,提出一种用于被遮挡特征学习的生成对抗网络(generative adversarial networks for learning occluded features,GANLOF).被遮挡特征学习网络分为被遮挡特征生成器、鉴别器两个部分.首先对数据集生成随机遮挡,作为模型的输入;然后利用生成器恢复被遮挡图片的池化特征,通过鉴别器区分恢复后的被遮挡池化特征与无遮挡图片池化特征,同时使用检测损失监督生成器,使恢复的被遮挡特征更准确.所提出被遮挡特征学习网络可以作为组件插入到任意的两阶段检测网络中.与Faster RCNN等已有模型相比,所提出模型在PASCAL VOC2007和KITTI数据集上的mAP (mean average precision)指标均有不同程度的提升.

Object detection is a fundamental task in computer vision. There often exist occlusions between objects in real life, which result in that some features of an object are missing, and detection accuracy is reduced. Therefore, we propose a generative adversarial network for learning occluded features(GANLOF). It is divided into two parts: the generator of occluded features and the discriminator. Firstly, we generate random occlusions for pictures in datasets, and the occluded pictures are the inputs of the network. Then we use the generator to restore pooling features in occluded regions, and the occluded pooling features and the un-occluded image pooling features are distinguished by the discriminator. Meanwhile,we use the detection loss to supervise the generator, so that the recovered occluded features are more accurate. The proposed GANLOF can be used as a component added into two-phase object detection networks. Compared with the Faster RCNN and other models, the mean average precision(mAP) of model is improved on the PASCAL VOC2007 dataset and the KITTI dataset.