一种基于区分区域定位的细粒度图像识别方法

A Fine Grain Image Recognition Method Based on Distinguishable Region Location

细粒度图像识别的目标为区分大类对象中的子类对象,由于子类对象间差别细微,使得细粒度图像识别较为困难。为此,提出一种基于区分区域定位的细粒度图像识别方法。首先由贝叶斯个性化排序损失(Bayesian Personalized Ranking Loss,BPRLoss)监督区域提议网络提议一些重要的局部区域,随后采用引入高效通道注意力模块的特征提取器提取局部区域的细粒度特征进行识别。同时采用标签平滑策略使同类靠近,不同类远离以监督网络学习对象有区别的特征,进一步促进网络定位区分区域。实验结果表明,所提方法在三种通用的细粒度图像识别数据集CUB-200-2011、FGVC Aircraft、Stanford Cars上取得了较高的识别准确率,分别为89.0%、93.9%、94.3%,相比导航网络(NTS-Net)有显著提升,分别提升1.5百分点、2.5百分点和0.4百分点。同时,所提方法较NTS-Net能够更为有效地定位区分区域和提取图像的细粒度特征。

The goal of fine-grained image recognition is to distinguish sub-class objects in large class objects.Because of the subtle differences between sub-class objects,fine-grained image recognition is more difficult.For this reason,a fine-grained image recognition method based on differentiated region location is proposed.Firstly,the Bayesian Personalized Ranking Loss(BPRLoss)supervised region proposes that the network proposes some important local regions,and then uses the feature extractor introducing the efficient channel attention module to extract the fine-grained features of the local regions for recognition.At the same time,the tag smoothing strategy is used to make the same class close and different classes far away to monitor the different characteristics of the network learning objects,and further promote the network location to distinguish regions.The experimental results show that the proposed method has achieved high recognition accuracy on three common fine-grained image recognition data sets CUB-200-2011,FGVC Aircraft and Stanford Cars,which are 89.0%,93.9%and 94.3%,respectively.Compared with the navigation network(NTS-Net),it has significantly improved by 1.5 percentage points,2.5 percentage points and 0.4 percentage points respectively.At the same time,the proposed method is more effective than NTS-Net in locating and distinguishing regions and extracting fine-grained features of images.