No abstract

Abstract. Introducing Deep Learning has been successful in improving the performance of automated writer identification systems. However, using very large patch sizes as input to CNN consumes a lot of machine resources and requires a lot of training time. To overcome this problem, many researchers use resized images.In this paper, we will try to make a comparative study between several patches sizes which were then resized to a normalized size of 32 × 32. Our aim is to elaborate the best recommendations for choosing the image resizing in order to increase the CNN performance. Thus, we will carry our tests on three databases. The first is CVL, a Latin dataset with 310 writers, the second is CERUG-CH a Chinese dataset with 105 writers and the last is KHATT that contains the Arabic writings of 1000 writers. To see if the type of CNN model impacts the results conducted on resized images, we deploy two models: ResNet-18 and MobileNet. The main finding is that the best results correspond to the resizing values of the images which makes it possible to have the average line height of the writings closer to the height of the CNN patches.

Writer Identification has gained increasing importance in the scientific community in recent years. In this paper, we propose an approach based on the combination of local textural descriptors and encoding methods (VLAD and Triangulation Embedding). The tests carried out in the bilingual LAMIS dataset made it possible to reach 100% in the Arabic version and 100% in the French version.

The use of computers and automatic systems has enabled scientific researchers to improve the classification rate in the field of writer identification. In our paper, we will propose an identification system based on the use of Histogram of Gradient Angle Distribution (HGAD) in square patches centered around Harris Keypoint locations. A global descriptor per image is calculated subsequently via the VLAD encoding of the local descriptors relating to the histograms of the square patches. The study carried out on two public datasets CVL and BFL made it possible to achieve very interesting identification rates with 99.4% in BFL and 99.7% in CVL.