Spatially-Consistent Feature Matching and Learning for Heritage Image Analysis

Progress in the digitization of cultural assets leads to online databases that become too large for a human analyze. Moreover, some analyses might be challenging, even experts. In this paper, we explore two applications computer vision analyze historical data: watermark recognition and one-shot repeated pattern detection artwork collections. Both problems present challenges which believe representative ones encountered heritage applications: limited supervision is available, tasks are fine-grained recognition, data comes several different modalities. also highly practical, as recognizing watermarks makes it possible date locate documents, while detecting patterns allows exploring visual links between artworks. We demonstrate on both benefits relying deep mid-level features. More precisely, define an image similarity score based geometric verification features show how spatial consistency can used fine-tune out-of-the-box target dataset with weak or no supervision. This paper relates extends our previous works (Shen et al. Discovering art collections spatially-consistent feature learning, 2019; Shen Large-scale new consistency-based approach, 2020). Our code available at http://imagine.enpc.fr/~shenx/HisImgAnalysis/ .