Misako Suwa

Segmentation of connected handwritten numerals by graph representation

This paper proposes a new algorithm for separating a touching pair of digits by using the graph-representation of the pattern. The segmentation can be regarded as grouping these edges and vertices into two disconnected sub-graphs. This process is executed by applying graph theory methods and certain heuristic rules. Since the boundaries of patterns are determined along the edges, the shapes of the segmented digits can be restored with high quality. The algorithm can segment not only simply connected cases but also multiply connected ones. The results of the performance evaluation using the NIST database are also presented.

Handwritten Chinese address recognition

A handwritten Chinese address recognition (HCAR) system is proposed in this paper. Handwritten Chinese address recognition is a difficult problem. Handwritten Chinese characters are characterized by large vocabulary, complicate structure, irregular distortion and touching characters etc. The proposed approach takes good advantage of Chinese address knowledge, and applies key character extraction and holistic word matching to solving the problem. Different from conventional approach, proposed approach can avoid the character segmentation error successfully. Experimental results show the proposed approach is very effective.

Segmentation of handwritten numerals by graph representation

A new algorithm is proposed for segmenting simply and multiply connected digits. It also removes ligatures. After thinning the pattern, the edges and vertices are extracted and the pattern is represented as a connected graph. Then the matrices relating to the graph are calculated. To determine the segmentation path, both graph theory techniques and heuristic rules are used. The boundaries of digits are calculated to make the width of touching strokes uniform. The separated digits thus have a more natural shape than can be achieved using algorithms that split patterns using straight lines or line segments.

Handwritten Numeral Recognition Using Personal Handwriting Characteristics Based On Clustering Method

To improve recognition rate, it is important not only to utilize one character feature but personal handwriting characteristics. This paper realizes above approach based on our investigation result that characters written by the same writer have similar shapes and that there are several shapes even in the same category. In our method, clustering method is used to absorb the variance of character shapes in the category. First, character recognition for each character is executed. Next, misrecognized character candidates are extracted as isolated cluster by within-category clustering. Then, recognition results of the extracted characters are amended by between-category clustering which evaluates the distance between the cluster composed of misrecognized characters and the cluster composed of correctly recognized characters in every categories. Finally, experimental results shows that recognition rate is remarkably improved by our method.

Separation of overlapped color planes for document images

Color plane separation is very useful in processing color document images. Many reported methods take it as a multi-class classification problem and work not well in overlapped color regions. This paper proposed a simple but effective linear projection based method for separating overlapped color planes. The separation task is taken as a probability problem, i.e., in the output plane, target color should have high response and the other colors should have low response, or vice versa. Furthermore, it assumes that the number of foreground colors is low, typically one to four, and overlapped areas contain mixed colors instead of opaque covering. Experimental results demonstrate the effectiveness and flexibility of our method.

Occluded text restoration and recognition

Text occlusion is among the most intractable obstacles for OCR engines. A typical example in document images is visible watermark characters, which are often occluded by foreground contents. This paper proposes a solution by restoring watermark characters before recognition. The text restoration process consists a core module as patch-based restoration method, which reconstructs the missing areas by referring to similar patches from undamaged areas. The filling sequence is in a order based on the structure complexity inside each patch, which helps to suppress reconstruction error propagation. Furthermore, the patch size is adaptively selected based on the local character stroke width. Experiments show that the proposed method produces good restoration quality and effectively improves the recognition rate of the following OCR process. Furthermore, the algorithm is optimized based on statistical analysis model and the processing time meets the real-time responding requirement.

Accuracy improvement for handwritten Japanese word recognition by combination of character and word recognizer

This paper proposes a combination method of character recognizer (CR) and word recognizer (WR) for Japanese handwritten words. CR is composed of character segmentation, isolated character recognition, and postprocessing. On the other hand, WR recognizes word images holistically based on a given word lexicon. These methods have been used as key modules in handwritten address recognition but have not been compared quantitatively up to now. In this paper, CR and WR are first empirically compared with each other using synthesized words generated from character images. Based on the comparison, a combination method is then proposed Experimental results show that the combination approach can achieve better recognition performance than the other two methods.

Handwritten Chinese Address Recognition Based on Enhanced Key Character Extraction and Holistic Word Matching

A handwritten Chinese address recognition (HCAR) system is proposed in this paper. Handwritten Chinese addresses are difficult to recognize because handwritten Chinese characters are characterized by a large vocabulary, complicated structure, irregular distortion and touching characters, etc. The proposed approach takes good advantage of Chinese address knowledge, and applies key character extraction and holistic word matching in solving the problem. Different from conventional approaches, the proposed approach can successfully avoid character segmentation errors. Experimental results show the proposed approach is very effective.

Improvement of numeral recognition using personal handwriting characteristics based on clustering

Correctly recognizing characters with peculiarities for each writer is a difficult problem. The process of absorbing variations in individual writing by creating an individual dictionary is also difficult when a writer is not specified and the total number of writers is large. In this paper the authors propose a method to improve the results of isolated character recognition in forms in which the same writer writes many characters by taking the characteristics of the writers writing on a form as a character distribution in a character feature space. In concrete terms, the authors first perform isolated character recognition on all characters on the same form. Then, based on the results of isolated character recognition, clustering of input character groups is performed for each character category. Clusters which are very likely to include misrecognized characters from isolated character recognition are extracted based on the results of clustering. Then character categories in the extracted cluster are automatically amended based on the distance from all clusters in other categories. In the same fashion, automatic amending is performed for rejected characters. Based on experiments to evaluate handwritten numerals on OCR forms, the authors show that the precision of numeral recognition is improved by using this approach as a form of postprocessing for isolated character recognition.

Color-Mixing Correction of Overlapped Colors in Scanner Images

Color mixing occurs between the colors of ink and paper when the former is non-opaque. For non-white paper, even the hue of a figure in such an ink may change. In such cases, color conversion is required to extract only the pattern of a specified ink-color from a document image. In this paper, we consider an effective model of the Yule-Nielsen type with the Beer-Lambert law for scanner images of handwritten forms with ballpoint pens, and propose a new method of correcting color mixing. These parameters are calculated using only the information obtained from a pair of scanner images with different-colored paper and the same ink color. The experimental results confirm the feasibility of this approach.