Article Preview
Top1. Introduction
The ever-growing evolution in document digitization technologies serves several challenges one of which is handwriting recognition and word spotting. Word spotting consists of retrieving information from huge scanned document databases. The volume and complexity of such databases are making this task impossible manually (time-efficiency and cost-effectiveness). Many researchers around the world have tackled this challenge, and some of them are working on automated word-spotting systems. These systems are capable of extracting relevant and precise information from large and complex document databases. Content-Based Image Retrieval (CBIR) is an image retrieval technique (not specific to documents) used to retrieve information from large image databases (Ciciani, 1995). This technique relies on computing similarities between visual features (textures, colors, and shapes). CBIR is handy when it comes to performing a global search in a natural image, but becomes unfortunately very limited in case of a partial search (within the documents themselves). Therefore, other types of systems such as Document Image Retrieval Systems (DIRS) are more suitable for this type of images. They are based on two categories of techniques: techniques based on recognition (manual or automatic) and techniques based on word spotting. The recognition-based techniques are very accurate for recently written (modern handwriting), and good quality printed or handwritten documents, but lack effectiveness for old and degraded documents. Moreover, these techniques are often limited to a single language. Consequently, a handwriting recognition system for ancient documents or documents written in a rare language is very complicated to design and validate.
Afterwards, techniques of word spotting destined for documents have been proposed to overcome the problems resulting from recognition-based techniques. Our interest is for the word spotting techniques related to the interest point detection, description and matching on old and degraded handwritten document images.
The keypoints detectors and descriptors like SIFT (Lowe, 2004) and SURF (Bay, 2006) are widely used in various CBIR and DIRS applications due to their invariance to image scale and rotation and their robustness relatively efficient compared to older algorithms (corner, blob or region detectors). Several variants have been proposed in the literature (see section 2) to improve or adapt these methods to the specificities of certain application problems encountered. For each application, the performance of keypoint detection and description is compared for scale change, rotation change, blur change, illumination change, and/or affine transformations. The major drawback of these approaches (for detection or description phase) lies in the fact that the invariance to the change of illumination and colors is not taken into account (or taken into account in a partial and insufficient way in some variants). This disadvantage makes it difficult to apply these approaches to old and degraded handwritten documents with a very variable writing ink density (due to the use of the pen). For example, the SURF detector uses a fixed empirical threshold for keypoint filtering that is unrelated to the present local variations. This is a significant drawback on degraded handwritten, which requires the use of adaptive or automatic thresholding (global or local) to achieve better robustness to the degradation, noise, blur, illumination, ink or color change. These local variations destabilizes the keypoint detection and description because of the increase of their rates of under or over-detection. On the handwriting, the invariance to the rotation does not have to take integrally, because the rotation is partially vital to distinguish some pen movement. For the matching step, many literature techniques are fixed to matches descriptors vectors with ignoring the spatial relationship (Cartesian) that can excite between two matched keypoints. We trust that this information is significant and must be taken into account (see section 2). This paper addresses the problem of these invariances, their impact on the three phases of the word-spotting system (keypoints detection, description, and matching) and the importance of taking into account of the spatial relationships at the matching stage.