毕业生知识库

        甲骨文可以追溯到中国的商朝，是世界上最古老的象形文字之一。甲骨文字的识别对于考古学、古文字学以及历史年代学有着重要的意义。甲骨文字图像分为临摹甲骨文字图像与拓片甲骨文字图像两类，临摹甲骨文字图像为拓片甲骨文字图像经过专家处理后得到的高清图像，修复了拓片甲骨文字图像的残缺、噪声严重问题。本文针对上述两类甲骨文字图像的识别方法进行了研究，主要工作内容分为两部分：

       提出了一种基于深度度量学习的临摹甲骨文识别算法，有效克服了临摹甲骨文字样本量不足、样本类别不均衡以及样本类内差异过大的问题。首先基于卷积神经网络（Convolutional Nerual Network, CNN）将临摹甲骨文字图像映射到一个特征空间，使得特征间的欧式距离可以衡量对应的临摹甲骨文字图像间的差异。继而在该空间内基于最近邻（Nearest Neighbor，NN）法则进行识别。本文提出的算法在已知类分类任务上验证了优于现有算法，同时在未知类别拒识、开放集识别等多个任务上验证了算法的有效性。最后，本文提出了一种原型裁剪算法，有效地缓解了最近邻分类速度慢的问题。

      建立了拓片甲骨文字识别的一个新框架，通过对临摹甲骨文字图像经过深度度量学习得到的特征的有效利用，明显的提升了模型的识别效果。首先，基于域自适应的方式将拓片甲骨文字图像映射到与临摹甲骨文字图像相同的特征空间并使得两者具有相似的分布；接着，通过度量学习的方式对拓片甲骨文字的特征进行调整——增大与异类临摹甲骨文字特征的距离、减小与同类临摹甲骨文字特征的距离；最后，基于最近邻准则对拓片甲骨文字进行识别。实验结果表明，本文提出的算法在拓片甲骨文识别任务上相对于现有算法有明显的提升，同时也实现了拓片甲骨文字的新类别识别。

    Oracle character is one kind of the earliest hieroglyphics, which can be dated back to Shang Dynasty in China. Oracle character recognition is important for modern archaeology, ancient text understanding, historical chronology, etc. There are two kinds of oracle character images, one is Linmo Oracle Character (LOC) images and the other is Tapian Oracle Character (TOC) images. TOC images are directly collected from oracle bones, after image denoising and restoration, it becomes LOC images. Our efforts and contributions are divided into two parts:

    To overcome the limitation and class imbalance of training data in LOC recognition, we propose a classi?cation method based on deep metric learning. A convolutional neural network (CNN) is used to map the LOC images to a feature space where the distance between different samples can measure their similarities such that classi?cation can be performed by the Nearest Neighbor (NN) rule. Experimental results show that the proposed method exceeds the existing ones in LOC recognition.

    For the recognition of TOC, a framework is proposed to improve the performance by taking advantage of LOC. We also use a CNN to map TOC to a feature space whose dimension is same as the one of LOC. Then, unsupervised domain adaptation is carried out in that feature space to make TOC have the same distribution with LOC. Moreover, the feature of TOC is further adjusted via deep metric learning in order to fulfill NN classification. Our proposed method not only achieves better performance in TOC recognition but also can figure out new categories.