一种结合非局部和多区域注意力机制的细粒度图像识别方法

细粒度图像识别的目标是对细粒度级别的物体子类进行分类,由于不同子类间的差异非常细微,使得细粒度图像识别具有非常大的挑战性。目前细粒度图像识别算法的难度在于如何定位细粒度目标中具有分辨性的部位以及如何更好地提取细粒度级别的细微特征。为此,提出了一种结合非局部和多区域注意力机制的细粒度识别方法。Navigator只利用图像标签便可以较好地定位到一些鉴别性区域,通过融合全局特征以及鉴别性区域特征取得了不错的分类结果。然而,Navigator仍存在缺陷:1)Navigator未考虑不同位置间的联系,因此所提算法通过引入非局部模块与Navigator相结合,来加强模型的全局信息感知能力;2)针对非局部模块未建立特征通道间联系的缺陷,构建基于通道注意力机制的特征提取网络,使得网络关注更加重要的特征通道。最后,所提算法在3个公开的细粒度图像库CUB-200-2011,Stanford Cars和FGVC Aircraft上分别达到了88.1%,94.3%,92.0%的识别精度,并且相比Navigator有明显的精度提升。

Fine-grained Image Recognition Method Combining with Non-local and Multi-region Attention Mechanism

The goal of fine-grained image recognition is to classify object subclasses at a fine-grained level.Because the differences between different subclasses are very subtle,fine-grained image recognition is very challenging.At present,the difficulty of this kind of algorithm is how to locate the distinguishable parts of fine-grained targets and how to extract fine-grained features of fine-grained levels.To this end,a fine-grained recognition method combining Non-local and multi-regional attention mechanisms is proposed.Navigator only uses image labels to locate some discriminative regions,and achieves good classification results by fusing global features and discriminative regional features.However,Navigator is still flawed.Firstly,the navigator does not consider the relationship between different locations,so the algorithm proposed in this paper combines the non-local module with the navigator to enhance the global information perception ability of the model.Secondly,aiming at the defect that the Non-local module does not establish the relationship between feature channels,a feature extraction network based on channel attention mechanism is constructed,which makes the network pay more attention to the important feature channels.Finally,the algorithm proposed in this paper achieves recognition accuracy of 88.1%,94.3%and 91.8%on three open fine-grained image databases,CUB-200-2011,Stanford Cars and FGVC Aircraft respectively,and has a significant improvement over Navigator.