Contrastive Dissimilarity for Writer Identification: a proof of concept

Writer identification is the task of identifying the author of a handwritten text, which is very important for forensic analysis and related areas. The difficulty grows as the number of potential writers increases and as handwriting styles often overlap. Real-world applications often face additional challenges, such as imbalanced datasets, where certain writers have more samples than others. This paper investigates the application of Contrastive Dissimilarity for writer identification, a method that addresses the problem of limited data and class imbalance. We focus on understanding how the Document Filter affects performance and how the approach performs under unbalanced conditions using two well-known datasets: IAM and CVL. By introducing artificial class imbalance into both datasets, we found that the Contrastive Dissimilarity Vector maintains comparable performance to its balanced counterpart, suggesting that it is capable of generating accurate representations even when data is limited or imbalanced.