随着近年来漫画逐步转向移动阅读方式，数字技术下漫画内容的理解迎来新的挑战。漫画人物作为漫画内容的关键因素，漫画人物识别的研究具有重要意义。然而，漫画图像大多由抽象线条组成，漫画人物会根据场景不同而改变大小、视角、表情或姿态，同时还要考虑图像中缩放、旋转、仿射变换、遮挡以及背景干扰等问题，这些问题都大大增加了人物识别工作的难度。 本文基于模板匹配算法，利用具有稳定性的SIFT特征，进行漫画集图像中的特定人物识别。具体地，首先对特定人物的模板图像和待识别的漫画图像提取SIFT特征，通过训练筛选出模板图像中对于人物识别更具判别力的SIFT特征；然后进行模板和漫画图像之间的SIFT匹配；利用拓扑聚类去除错误匹配，提升匹配效果；最后利用仿射变换将模板边框投影到漫画图像上，定位特定人物位置。 在包含100页的漫画数据集上进行实验，其中60页为训练集，40页为测试集。结果表明，本文提出的漫画人物识别方法具有可行性和有效性，识别特定人物的准确率可以达到70%以上。

Nowadays, more and more people read comics on mobile devices rather than read printed comic books, which brings on some new challenges to the analysis of comic content using digital techniques. As key parts of comic, comic characters play an important role in comic analysis, and their recognition draws significant research interest. However, most comic images are expressed by abstract line drawings, and comic characters are represented by various scales, viewpoints, expressions or poses in different scenes of comic. Besides, some other problems such as scaling, rotation, affine transformation, clutter and occlusion also should be considered, all of which make the recognition more difficult. This paper proposes a specific comic character recognition method, which addresses above challenges using SIFT feature-based template matching. Specifically, the first step is to extract SIFT features from template image containing a specific character and comic image respectively, and select some more discriminative SIFT features for characterizing characters in template image based on a training dataset. Secondly, SIFT matching is performed between template image and comic image. Thirdly, mismatches are discarded by using topological clustering, which can improve matching accuracy. Eventually, the character bounding box in template image can be projected on comic image using affine transformation and then the specific comic character can be localized. Experiments are performed on a dataset consisting of 100 comic images, in which 60 images are applied as training set and 40 images are applied as testing set. The results achieve recognition rate of 70%, which demonstrates the feasibility and effectiveness of the proposed method.