A Siamese Based One Shot Learning Network with a Watermark Enhancement Technique for Historical Watermark Recognition

Historical watermark identification is of paramount importance to archivists and historians for undertaking studies on historical documents. The identification problem is usually treated as an image processing task since most archives and documents are preserved in pictorial format. However, because of the large variety of watermarks, scarcity of clean samples for new watermarks, crowded and noisy samples, diverse forms of representations, both modest distinctions between classes and substantial intra-class variations, these watermarks are often difficult to recognize. This paper proposes a one-shot classification pipeline for historical watermark identification using only one example of each class. The proposed pipeline is designed based on the Siamese network architecture where an EfficientNetB0 backbone is employed as the feature extractor. Due to the use of Siamese based proposed architecture, a contrasting loss function can be formulated to update the network parameters, which enables the network to learn to distinguish between similar and dissimilar pairs of watermark images more efficiently. A thresholding-based image pre-processing technique is also proposed for enhancing the watermarks before feeding them to the feature extractor. Experimental results on a publicly available large-scale historical watermark dataset reveal that the proposed method can lead to better performance than some one-shot classification methods reported in the literature.