基于语义信息跨层特征融合的细粒度鸟类识别

FINE-GRAINED BIRD RECOGNITION BASED ON SEMANTIC INFORMATION CROSS-LAYER FEATURES FUSION

有效识别各种鸟类目标具有重要的生态环境保护意义。针对不同种类鸟类之间差别细微、识别难度大等问题,提出一种基于语义信息跨层特征融合的细粒度鸟类识别模型。该模型由区域定位网络、特征提取网络和一种跨层特征融合网络(Cross-layer Feature Fusion Network,CFF-Net)组成。区域定位网络在没有局部语义标注的情况下,自动定位出局部有效信息区域;特征提取网络提取局部区域图像特征和全局图像特征;CFF-Net对多个局部和全局特征进行融合,提高最终分类性能。结果表明,该方法在Caltech-UCSD Birds200-2011(CUB200-2011)鸟类公共数据集上,取得了87.8%的分类准确率,高于目前主流的细粒度鸟类识别算法,表现出优异的分类性能。

In view of the subtle differences between different bird species and the difficulty of recognition,we propose a fine-grained bird recognition model based on cross-layer feature fusion of semantic information.It consists of regional location network,feature extraction network and cross-layer feature fusion network(CFF-Net).The regional location network automatically located the local effective information region without local semantic annotation;feature extraction network extracted local and global image features;CFF-Net combined multiple local and global features to improve the final classification performance.The results show that the classification accuracy is 87.8%on Caltech-UCSD Birds200-2011(CUB200-2011)dataset,which is higher than the current mainstream fine-grained bird recognition algorithm.And it shows excellent classification performance.