Exploring Feature Extraction Methods for Unsupervised Symbol Recognition in Historical Ciphers

Sammanfattning

Historical ciphers are difficult to transcribe automatically because they often contain rare or invented symbol alphabets and only limited labelled data. This thesis investigates how different visual feature extraction methods affect unsupervised symbol clustering for historical cipher transcription. The study compares handcrafted descriptors such as SIFT and HOG, CNN-based features, hybrid feature combinations, and an entropy-based feature weighting method on the BNF cipher dataset. The extracted features are clustered using k-means and evaluated with Adjusted Rand Index, Normalized Mutual Information, and Character Error Rate. The results show that HOG achieves the best overall performance across all tested settings, outperforming both SIFT and the CNN-based representations. VGG16 with ImageNet pre-training performs competitively but does not surpass HOG, while SIFT performs substantially worse. Fine-tuning VGG16 on Omniglot, combining SIFT and VGG16 features, and applying entropy-based feature weighting do not consistently improve over the strongest individual methods. The findings suggest that a relatively simple gradient-based descriptor can capture cipher symbol shapes more effectively than deeper high-dimensional representations for the BNF dataset. The thesis also shows that clustering metrics and transcription-oriented error rates capture different aspects of performance, and that both should be considered when evaluating unsupervised cipher symbol recognition.