A Comparison of CNN and Conventional Descriptors for Word Spotting Approach: Application to Handwritten Document Image Retrieval

Abstract

Natural images are easier to represent in feature space than textual images due to the reduced complexity and thus do not require greater learning capacity. The visual information representation is an important step in content-based image retrieval (CBIR) systems, used for searching relevant visual information in large image datasets. The extraction of discriminant features can be carried out using two approaches. Manual (conventional CBIR), based on preselected features (colors, shapes, or textures) and automatic (modern CBIR) based on auto-extracted features using deep learning models. This second approach is more robust to the complexity relative to textual images, which require a deep representation reaching the semantics of text in the image. DIRS (Document Image Retrieval Systems) are CBIR systems related to documents images that propose a set of efficient word-spotting techniques, such as the interest points based techniques, which offer an effective local image representation. This paper presents an overview of existing word retrieval techniques and a comparison of our two proposed word-spotting approaches (interest points and CNN description), applied on handwritten documents. The results obtained on degraded and old Bentham datasets are compared with those of the literature.