联合语义部件的鸟类图像细粒度识别  

Joint Semantic Parts for Fine-Grained Bird Images Recognition

作　　者：赵毅力[1,2] 徐丹 Zhao Yili;Xu Dan(School of Information,Yunnan University,Kunming 650091;School of Big Data and Intelligent Engineering,Southwest Forestry University,Kunming 650224)

机构地区：[1]云南大学信息学院,昆明650091 [2]西南林业大学大数据与智能工程学院,昆明650224

出　　处：《计算机辅助设计与图形学学报》2018年第8期1522-1529,共8页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(61662072;61540062);云南省教育厅基金(2015Y285;2016CYH03);云南省应用基础研究项目(2014FA021)

摘　　要：由于子类别的高度相似性引起的类间微小差异,以及姿态、尺度和旋转方面的类内变化,使得细粒度图像识别成为一个具有挑战性的计算机视觉问题.为了对鸟类图像进行细粒度识别,提出一种联合语义部件的深度卷积神经网络模型.该模型由2个子网络组成:一个是语义部件检测子网,使用深度残差网络对鸟类图像语义部件进行精确定位;另一个是分类子网,使用三路深度残差网络对检测子网检测到的语义部件进行联合分类.收集了一个新的鸟类图像数据集YUB-200-2017,用于鸟类图像细粒度识别实验.结果表明,在YUB-200-2017和CUB-200-2011数据集上,文中方法具有较高的语义部件检测精度和识别准确率.Fine-grained image recognition is a challenging computer vision problem, due to small inter-classvariations caused by highly similar subordinate categories, and the large intra-class variations in poses,scales and rotations. In order to perform fine-grained recognition on bird images, this paper proposes a deepconvolution neural networks model collaborated with semantic parts detection. The model consists of twomodules, one module is a parts detector network, and another module is a three-stream classification networkbased on deep residual network. In the meantime, a new bird images dataset was collected and labeled to facilitythe research of fine-grained bird images recognition. Experiment results on YUB-200-2017 andCUB-200-2011 illustrate the proposed model has higher part detection and image classification accuracycomparing with state-of-the-arts fine-grained bird image recognition approaches.

关 键 词：细粒度图像识别 语义部件检测 深度学习 卷积神经网络 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]