Abstract
This paper deals with historical encrypted manuscripts and introduces an automated method for the detection and transcription of ciphertext symbols for subsequent cryptanalysis. Our database contains documents used in the past by aristocratic families living in the territory of Slovakia. They are encrypted using a nomenclator which is a specific type of substitution cipher. In our case, the nomenclator uses digits as ciphertext symbols. We have proposed a method for the detection, classification, and transcription of handwritten digits from the original documents. Our method is based on Mask R-CNN which is a deep convolutional neural network for instance segmentation. Mask R-CNN was trained on a manually collected database of digit annotations. We employ a specific strategy where the input image is first divided into small blocks. The image blocks are then passed to Mask R-CNN to obtain detections. This way we avoid problems related to the detection of a large number of small dense objects in a high-resolution image. Experiments have shown promising detection performance for all digit types with minimum false detections.