基于角色多语义特征识别的动漫版权认证技术研究

目前，我国动漫作品在司法保护实践中面临侵权认定举证困难等困境，因此有效地提取动漫声视觉特征，实现版权保护成为亟待解决的问题。本研究针对上述问题，依次提取动漫角色肖像、语言语音、个性动作等多语义特征进行版权认证。在漫画特征提取方面，针对肖像，在混合域下用四元数图像深度分解的方式提取阶跃边缘特征；针对语言，采用语言风格数字化建模的方式提取统计特征；针对个性动作，采用运动模板的方式提取运动角度特征。在动画特征提取方面，针对肖像，在混合域下用运动补偿深度分解的方式提取阶跃边缘特征；针对语音，采用音频信息融合的方式提取感知、方向、能量分布特征；针对个性动作，采用金字塔LK光流法提取运动向量特征，并结合感知哈希、相似哈希、主成分及线性判别分析等技术实现特征的降维及完全互补码、视觉密码等技术提高安全性。该方案为动漫版权认证提供依据，并有效解决特征构造和版权识别中安全性低、特征构造难、准确率低等问题。

At present, China's anime works are facing difficulties such as infringement certification in judicial protection practice. Therefore, it is an urgent issue to effectively extract the sound and visual characteristics of anime and realize copyright protection. In this study, aiming at the above problems, multi-semantic features such as portraits, Speech language, and individual actions of anime characters were sequentially extracted for copyright authentication. In the aspect of comic feature extraction, for portraits, step edge features are extracted domain by using quaternion image depth decomposition in the mixed domain; For language, statistical features are extracted by using digital modeling of language style; For individual actions, the motion angle feature is extracted by using motion template. In the aspect of animation feature extraction, for portraits, step edge features are extracted by using motion-compensated depth decomposition in the mixed domain; For speech, the features of perception, direction, and energy distribution are extracted by using audio information fusion; For individual actions, motion vector features are extracted by using pyramid Lucas-Kanande optical flow method. At the same time, combines techniques such as perceptual hashing, similar hashing, principal components, and linear discriminant analysis to achieve feature reduction. And combines technologies such as complete complementary code and visual cryptography to improve security. This solution can provide basis for anime copyright certification, and effectively solve the problems of low security, difficult feature construction, and low accuracy in feature construction and copyright recognition.