基于卷积神经网络语义检测的细粒度鸟类识别

细粒度识别的主要目的是在相同基本类别下对其繁多的子类别进行区分。不只局限于头和躯干的定位现状,提出了一种基于Faster RCNN联合语义提取和检测的分类方法。通过引入自上而下的方法来生成七个小语义部位,既大大减少了候选区域的个数,又提高了分类的效率。检测子网可以和区域候选生成网络(RPN)共享卷积特征,结果使得区域建议几乎不花时间,从而可以生成高质量并且具有局部特征的区域建议框,便于Fast RCNN的检测。相对于其他鸟类识别研究,实验中鸟类识别准确率达到了88.37%,提高了识别效率。说明联合语义的Faster RCNN网络适用于鸟类的细粒度识别。

Fine - grained Bird Recognition Based on Convolution Neural Network Semantic Detection

The main purpose of fine-grained identification is to distinguish between its many subcategories under the same basic categories. This paper not only confines the localization of head and torso, and proposes a classification method based on Faster RCNN joint semantic extraction and detection. By introducing the top-down method to generate seven small semantic parts, it greatly reduces the candidate region Number, but also improve the efficiency of classification.Since the detection subnet can share the convolution characteristics with the regional proposals network (RPN), the result is that the area is proposed to take almost no time, so that it can generate high quality and local characteristics of the regional proposal boxes to facilitate the detection of Fast RCNN. Compared with other bird recognition studies, the accuracy of birds recognition in the experiment reached 88.37%, which improved the recognition efficiency. The joint semantic Faster RCNN network is suitable for fine-grained recognition of birds.