基于视觉Transformer内在归纳优化的齐白石虾画真假鉴定
A Qi Baishi's Shrimp Paintings Identification Algorithm Based on Intrinsic Inductive Properties Optimized Visual Transformer
-
摘要: 当前书画艺术品市场赝品众多, 给书画艺术品收藏者带来了极大的经济风险, 并且严重扰乱书画艺术市场秩序. 针对书画艺术品真假鉴别的数据集一般较小的特点, 设计了数据高效的齐白石虾画自动真假鉴别的算法. 以视觉Transformer为基础架构, 通过对视觉Transformer的标记位置编码方式进行改进, 同时以跨架构表征知识蒸馏对模型进行训练, 改善视觉Transformer的内在归纳特性, 减少模型对训练数据的过度依赖, 有效地解决了齐白石虾画真假鉴别数据集较小的挑战. 分别在有429幅画的齐白石虾画真假鉴别数据集、有96 013幅画的WikiArt数据集和有42 479幅画的ArtDL数据集上获得的实验表明, 该方法有效地应对了齐白石虾画真假鉴别任务中数据集小的挑战, 并在齐白石虾画真假鉴别任务达到了优于其他方法的分类性能.Abstract: The current painting market is full of forgeries, which brings great economic risks to the collectors and disrupts the order of the painting market. Since the data set for the authenticity identification of calligraphy and painting artworks are generally small, this paper proposed a data-efficient algorithm for automatic authenticity identification of Qi Baishi shrimp paintings. The proposed method takes the visual transformer as the backbone and improves the token embedding strategy of the visual transformer with relative position embedding. Besides, this paper trained the model with cross-architecture representation knowledge distillation, which improves the inductive bias of the visual transformer, and reduces the model's demand for training data. The proposed method effectively improves the model's performance in a small dataset of Qi Baishi shrimp painting authenticity identification. Experiments performed on the authenticity identification dataset of Qi Baishi shrimp paintings with 429 paintings, the WikiArt dataset with 96 013 paintings, and the ArtDL dataset with 42 479 paintings show that the proposed method can effectively identify the authenticity of Qi Baishi shrimp paintings.