摘要
深度学习方法在图像内容检测中取得良好效果并得到广泛应用。自然场景中Logo(标识)图案的检测具有很强的商业和社会需求,但获取其适用于深度学习方法的训练数据却并不容易。为解决上述问题,针对自然场景图片中包含Logo的检测和识别,本文提出一种生成对应训练数据集的合成方法。根据简单的输入和参数设置,该方法能够自动生成大量带有特定Logo并且符合自然场景特征的图片以及相应标注数据。这些生成数据可用于自然场景Logo检测识别的训练数据集,降低了深度学习网络模型的训练成本。使用本文方法合成的数据集训练的深度网络模型,可在FlickrLogos-32标准测试数据集上达到63.9%的平均准确率(mAP),接近使用大量真实人工标注数据的效果,体现了本文方法的有效性。
Deep learning has been widely used in image recognition. While logo detection is essential in commercial and social applications, big training data that deep learning needs is difficult to obtain. A data synthesis method is proposed for generating training dataset of natural image for logo detection. Given a few raw data and optional parameters, this method can automatically generate massive training data comprising of synthesized images and respective annotations. Using the synthesized training data instead of hand-labeled ones will dramatically reduce the training cost. Experiment on FlickrLogos-32 reports a mAP of 63.9%,a competitive result against that of the model using hand-labeled real-scene data.
作者
欧啸天
胡伟
OU Xiao-tian;HU Wei(School of Information Science and Technology, Beijing University of Chemical Technology, Beijing 100029, China)
出处
《电子设计工程》
2018年第7期179-184,共6页
Electronic Design Engineering
