Jonghyon Yi

Segmentation for MRC compression

Mixed Raster Content (MRC) is a standard for efficient document compression which can dramatically improve the compression/quality tradeoff as compared to traditional lossy image compression algorithms. The key to MRCs performance is the separation of the document into foreground and background layers, represented as a binary mask. Typically, the foreground layer contains text colors, the background layer contains images and graphics, and the binary mask layer represents fine detail of text fonts. The resulting quality and compression ratio of a MRC document encoder is highly dependent on the segmentation algorithm used to compute the binary mask. In this paper, we propose a novel segmentation method based on the MRC standards (ITU-T T.44). The algorithm consists of two components: Cost Optimized Segmentation (COS) and Connected Component Classification (CCC). The COS algorithm is a blockwise segmentation algorithm formulated in a global cost optimization framework, while CCC is based on feature vector classification of connected components. In the experimental results, we show that the new algorithm achieves the same accuracy of text detection but with lower false detection of non-text features, as compared to state-of-the-art commercial MRC products. This results in high quality MRC encoded documents with fewer non-text artifacts, and lower bit rate.

Color Registration Error Reduction in Document Scanner Using Iterative Velocity Command Synthesis

Document scanners enjoy widespread use for converting hardcopy documents into digital images. Improving image quality is a key issue for scanner manufacturers. This paper proposes an ILC based step timing synthesis technique for reducing color registration error in stepper motor driven document scanners with no real-time feedback. To quantify color registration error, a line pattern image based technique is devised. Based on the measured color registration error signal, a single iteration convergent iterative learning control scheme is proposed to synthesize a stepper motor step timing command for the scan region. Reduction in color registration error is achieved by compensating for velocity fluctuations in the constant speed portion of the motion. Experimental results demonstrate the effectiveness of the proposed approach in reducing color registration error in scanned image.Copyright

A new approach to JBIG2 binary image compression

The JBIG2 binary image encoder dramatically improves compression ratios over previous encoders. The effectiveness of JBIG2 is largely due to its use of pattern matching techniques and symbol dictionaries for the representation of text. While dictionary design is critical to achieving high compression ratios, little research has been done in the optimization of dictionaries across stripes and pages. In this paper we propose a novel dynamic dictionary design that substantially improves JBIG2 compression ratios, particularly for multi-page documents. This dynamic dictionary updating scheme uses caching algorithms to more effciently manage the symbol dictionary memory. Results show that the new dynamic symbol caching technique outperforms the best previous dictionary construction schemes by between 13% and 46% for lossy compression when encoding multi-page documents. In addition, we propose a fast and low-complexity pattern matching algorithm that is robust to substitution errors and achieves high compression ratios.

A recognition of map using the geometric relations between lines and the structural information of objects

A map consists of many lines, and the process to recognize an object in map is one to group lines that are consistent with the structure of the object. This paper proposes a novel method for recognizing roads and districts in land register maps. The lines that constitute an object are searched in a reduced search space by defining the some relations between lines, and objects are recognized by using the structural information.

Improved B-spline image registration between exhale and inhale lung CT images based on intensity and gradient orientation information

Registration of lung CT images acquired at different respiratory phases is clinically relevant in many applications, such as follow-up analysis, lung function analysis based on mechanical elasticity, or pulmonary airflow analysis, etc. In order to find accurate and reliable transformation for registration, a proper choice of dissimilarity measure is important. Even though various intensity-based measures have been introduced for precise registration, the registration performance may be limited since they mainly take intensity values into account without effectively considering useful spatial information. In this paper, we attempt to improve the non-rigid registration accuracy between exhale and inhale CT images of the lung, by proposing a new dissimilarity measure based on gradient orientation representing the spatial information in addition to vessel-weighted intensity and normalized intensity information. Since it is necessary to develop non-rigid registration that can account for large lung deformations, the B-spline free-form deformation (FFD) is adopted as the transformation model. The experimental tests for six clinical datasets show that the proposed method provides more accurate registration results than competitive registration methods.

Document image enhancement algorithm for digital color copier

A typical document image to be copied generally consists of various types of elements, for example, texts, halftone images, line drawings, and continuous tone images printed on background. To enhance the copy quality of such a compound document image, one or more image enhancement processes can be applied. However, it is not suitable to apply an enhancement method, which is probably appropriate to a typical kind of document element, on the document image entirely. Instead, it is preferable to discriminate each document element and to apply adequate image enhancement methods on respective document elements. The proposed document image enhancement method is comprised of a segmentation phase and an enhancement phase. In the segmentation phase, it classifies each pixel of the input document image into text, continuous tone image, halftone image and background by using a state transition machine, pixel-based features and run-based features. In the enhancement phase, it applies different enhancement methods on respective document elements. In this way, the dark gray texts are converted into black texts, and the edges of texts and continuous tone images are emphasized, while prohibiting halftone image regions from inducing erroneously enhanced artifacts.

A Map Mosaicking Method Using Opportunistic Search Approach with a Blackboard Structure

Map mosaicking is to integrate two or more map images having a coincident area by computing the rotational angle, the vertical and horizontal distances a map image has to move to overlap the coincident area. A solution of the problem is represented as a point in the parameter space with three axes: one for the rotational angle and the others for the vertical and horizontal distances. We extract local features from each map image, match them to make feature pairs, and project the feature pairs onto the parameter space. Traditional approaches using parameter spaces have suffered from a huge search space and computing time, for they project all the feature pairs onto the parameter space and search solutions by iterative optimization methods. We propose a new method that can give a solution not projecting all the feature pairs onto the parameter space but search opportunistically in a Blackboard structure.