面向瓷器AI模型的“难例负采样”（Hard Negative Mining）训练增强数据集

1. 高精度AI鉴定模型训练：直接用于训练需要进行精细化区分的AI鉴定模型。使用该数据集训练的模型，在区分“南宋官窑”与“龙泉仿官”等高难度任务上，性能会远超使用常规数据训练的模型。 2. 模型“对抗性训练”：可用于提升模型的鲁棒性，使其在面对风格接近的仿品或迷惑性样本时，不易被“欺骗”。 3. 发掘关键区分特征：通过分析模型为了区分“难例”而激活的神经元，可以反过来帮助人类专家理解AI究竟是依据哪些细微的视觉特征来进行高难度判断的。1. High-precision AI authentication model training: Directly utilized for training AI authentication models requiring fine-grained differentiation. Models trained on this dataset will significantly outperform those trained on conventional datasets in high-difficulty tasks such as distinguishing between Official Kiln Porcelain of the Southern Song Dynasty and Longquan imitative Official Kiln Porcelain. 2. Adversarial training enhancement: This dataset can be employed to boost the robustness of AI models, making them less prone to being "deceived" when confronted with stylistically similar imitations or misleading samples. 3. Extraction of key discriminatory features: By analyzing the neurons activated by the model when differentiating "hard examples", this dataset enables human experts to reverse-engineer the subtle visual features that the AI relies on to execute high-difficulty authentication judgments.

搜集汇总

背景与挑战

背景概述

该数据集是深圳市大雅斋国际拍卖有限公司申请的面向瓷器AI模型的训练增强数据集，采用难例负采样技术构建。数据集包含高迷惑性样本对，专门用于训练AI模型识别瓷器细微差别，提升在高精度鉴定和对抗性训练中的性能。数据以CSV格式存储，通过计算藏品间视觉相似度来构建训练三元组。