Computer Vision News - August 2021

Because of this, they came up with the idea to generate a new additional dataset, DIDA, which can cope with a variety of image sizes. DIDA contains 25,000 samples of each digit from 0 to 9 . “ This also allows the community to apply different types of CNN models ,” Amir points out. “With the 28 x 28 MNIST dataset the network contains a maximum of three to four convolutional layers, which is a problem. The more convolutional layers you have, the more information you can take from your images. This is another important characteristic of our dataset. Now, recognition accuracy is around 96-97%.” DIGITNET is a sub-project of a Swedish big data project funded by the Knowledge Foundation in collaboration with ArkivDigital . ArkivDigital has collected and digitalized over 90 million historical Swedish handwritten document images from between the 13th and 21st centuries. It is keen to provide an application like DIGITNET for its customers to improve the search capabilities on its database. Out of the 75,000 document images Amir and Hüseyin used for DIDA, they have only currently got permission to publisharound15,000wholedocument images of historical handwritten birth records, but they hope to be able to publish more soon. These birth records were used for another paper and dataset: SHIBR – The Swedish Historical Birth Records . Within the next year, they are hoping to expand their dataset further. They are preparing character datasets to use for word spotting to understand and label the document images as fully as possible and to make searching the database faster and more powerful. “We are nowworking on a newproject, ” Hüseyin tells us. 30 Computer Vision Project