基于模式挖掘的细粒度图像分类研究
首发时间:2020-03-13
摘要:由于细粒度物体之间的差异非常细微且存在于图像的局部区域中,因此细粒度图像分类的关键是定位到具有判别力的物体部件。本文提出了一种基于模式挖掘的细粒度图像分类方法 UPM~(Unsupervised Part Mining),该方法能够无监督地定位物体部件,进而提升细粒度图像分类的性能。首先,利用模式挖掘从预训练卷积神经网络的激活特征图中挖掘模式。这些模式往往在物体的外观和空间结构上具有一致性。受此启发,本文采用聚类算法对这些模式进行聚类,以定位物体的判别力部件。最后,构建一个多分支的网络结构来融合全局特征、物体级别特征和局部级别特征,完成细粒度图像识别任务。UPM方法在不使用图像类别标签、物体标注框和部件标注点等监督信息的条件下,仅依靠预训练模型能精准定位到细粒度图像中有判别力的部件区域。实验表明,UPM方法不仅获得了较高的分类精度,而且减少了模型的训练时间。
关键词: 细粒度图像分类; 模式挖掘; 无监督式定位; 深度学习
For information in English, please click here
Research on Fine-grained Image Classification Based on Pattern Mining
Abstract:Since subtle visual differences among fine-grained subcategories only locate at local regions, part localization is a key issue for fine-grained image classification. This paper proposes a method based on pattern mining, named UPM (Unsupervised Part Mining), which can localize discriminative parts under a fully-unsupervised setting. Firstly, the pattern mining algorithm is utilized to mine visual patterns in the feature maps extracted from a pre-trained convolutional neural network model. Inspired by the fact that these relevant meaningful patterns typically hold appearance and spatial consistency, then the clustering algorithm is performed on mined patterns to obtain the discriminative parts. Finally, a multi-stream classification model is built, which aggregates the image-level, object-level and part-level features simultaneously for classification. The UPM approach does not need any supervision, but can accurately localize the discriminative parts in the fine-grained image only relying on pre-trained CNN model. Experiments show that the UPM approach not only achieves the competitive performance, but also reduces the training time of the model.
Keywords: Fine-grained image classification pattern mining unsupervised localization deep learning
基金:
引用
No.****
动态公开评议
共计0人参与
勘误表
基于模式挖掘的细粒度图像分类研究
评论
全部评论0/1000