残差网络下基于困难样本挖掘的目标检测

Object Detection Based on Hard Examples Mining Using Residual Network

为了提高图像目标的检测精度,提出一种在残差网络下设计基于困难样本挖掘的目标检测算法。首先阐述基于深度学习的目标检测算法,即超快速区域卷积神经网络(Faster R-CNN)的工作原理,分析该算法存在的不足与改进方式。在Faster R-CNN的基础上,为了使模型能提取更有效的深度卷积特征,选取网络更深的残差网络替换原始的ZF或VGG网络。为了使学习到的网络模型有更强的泛化能力,在网络训练过程中,利用困难样本更新网络参数,使网络训练更充分。在Pascal VOC2007、Pascal VOC2007+Pascal VOC2012和BIT这三个数据集中进行训练和测试,实验结果显示,融合了两种方法的Faster R-CNN在这三个数据集上的检测精度分别提升了3.5%、7.1%、6.4%,提升效果明显。

In order to detect objects more accurately in images, an object detection algorithm based on hard example mining and residual network is proposed, which takes faster regional convolutional neural network （Faster R-CNN） as a benchmark. The working principle of Faster R-CNN is described based on deep learning, and the shortcomings and improvement methods of the algorithm are analyzed. Specifically, a deeper residual network is adopted to replace the original ZF or VGG network to extract more effective deep convolution features. In order to enhance the generalization ability of the learning network model, the network parameters are updated with hard examples during training. The experimental results on Pascal VOC2007, Pascal VOC2007 ＋ Pascal VOC2012 and BIT show that compared with Faster R-CNN, the proposed method improves detection accuracy by 3.5%, 7.1%, 6.4%, respectively, on the above three datasets.