As the big difference between natural type and art type, recognizing visual objects from photos to art paintings, cartoon pictures, or sketches introduces a great challenge. Domain adaptation focuses on overcoming the differences between different fields. It is an effective technology to bridge the cross-domain discrepancy by transferable features, while the existing domain-adaptive methods all need target domain images of the same category as source domain images to reduce domain shifts, which leads to limitations on target domain images. To solve this problem, we constructed an end-to-end unsupervised model called adaptive depiction fusion network (ADFN). Compared with other domain adaption methods, ADFN recognizes visual objects in art works by using only their natural type. It reinforces adaptive instance normalization technology to embed the depiction offset into the source domain features. At the meantime, we also provide a complete benchmark, cross-depiction-net, which is large and various enough to overcome the lack of data for this problem. To properly evaluate the performance of the ADFN, we compared it to different state-of-the-art methods (DAN, DDC, Deep-coral, and MRAN) on cross-depiction-net dataset. The results show that our model is superior to the state-of-the-art methods.

中文翻译：

---

从自然类型中识别艺术作品图像：一种深度自适应描绘融合方法

作为自然类型和艺术类型之间的巨大差异，识别从照片到艺术绘画、卡通图片或素描的视觉对象带来了巨大的挑战。领域适配侧重于克服不同领域之间的差异。通过可迁移特征来弥合跨域差异是一种有效的技术，而现有的域自适应方法都需要与源域图像相同类别的目标域图像来减少域偏移，从而导致对目标域图像的限制。为了解决这个问题，我们构建了一个端到端的无监督模型，称为自适应描述融合网络（ADFN）。与其他领域适应方法相比，ADFN 仅使用其自然类型来识别艺术作品中的视觉对象。它加强了自适应实例归一化技术，将描述偏移嵌入到源域特征中。同时，我们还提供了一个完整的benchmark，cross-description-net，它的规模和多样性足以克服这个问题的数据不足。为了正确评估 ADFN 的性能，我们在交叉描述网络数据集上将其与不同的最新方法（DAN、DDC、Deep-coral 和 MRAN）进行了比较。结果表明，我们的模型优于最先进的方法。和 MRAN）在交叉描述网络数据集上。结果表明，我们的模型优于最先进的方法。和 MRAN）在交叉描述网络数据集上。结果表明，我们的模型优于最先进的方法。