Ehsanollah Kabir

In this study, a method for holistic recognition of handwritten Farsi words is proposed, which fuses the outputs of right‐to‐left (RtL) and left‐to‐right (LtR) hidden Markov models (HMMs). The experimental results on 16,000 images of 200 names of Iranian cities, from the ‘Iranshahr 3’ are presented and compared with those methods using only RtL or LtR models. Experimental results show that the main sources of error are similar beginnings or similar endings of the words. Since RtL and LtR models when dealing with the words behave differently, there is notable error diversity between the two classifiers in such a way that their combination increases the recognition rate. Compared to the RtL‐HMM, the product of output scores of the RtL and LtR‐HMMs reduces the classification error to about 6, 6 and 3%, for three different feature sets. A subjective error analysis on the results is also provided.

Despite the recent progress in OCR technologies, whole-book recognition, is still a challenging task, in particular in case of old and historical books, that the unknown font faces or low quality of paper and print contributes to the challenge. Therefore, pre-trained recognizers and generic methods do not usually perform up to required standards, and usually the performance degrades for larger scale recognition tasks, such as of a book. Such reportedly low error-rate methods turn out to require a great deal of manual correction. Generally, such methodologies do not make effective use of concepts such redundancy in whole-book recognition. In this work, we propose to train Long Short Term Memory (LSTM) networks on a minimal training set obtained from the book to be recognized. We show that clustering all the sub-words in the book, and using the sub-word cluster centers as the training set for the LSTM network, we can train models that outperform any identical network that is trained with randomly selected pages of the book. In our experiments, we also show that although the sub-word cluster centers are equivalent to about 8 pages of text for a 101- page book, a LSTM network trained on such a set performs competitively compared to an identical network that is trained on a set of 60 randomly selected pages of the book.

Text segmentation is a live research field with vast new areas to be explored. Separating text layer from graphics is a fundamental step to exploit text and graphics information. The language used in the map is a challenging issue in text layer separation problem. All current methods are proposed for non-Persian language maps. In Persian, text strings are composed of one or more subwords. Each subword is also composed of one to several letters connected together. Therefore, the components of the text strings in Persian are more diverse in terms of size and geometric form than in English. Thus, the overlapping of the Persian text and the lines usually produces a complex structure that the existing methods cannot handle with the necessary efficiency. For this purpose, the stroke width variety of the input map is calculated, and then the average line width of graphics is estimated by analyzing the content of stroke width. After finding the average width of graphical lines, we classify the complex structure into text and graphics in pixel level. We evaluate our method on some variety of full crossing text and graphics in Persian maps and show that some promising results in terms of precision and recall (above 80% and 90%, respectively) are obtained.

Historical printed books OCR is one of the challenging tasks in the area of document image analysis. Low quality of print and paper and unfamiliar font faces are the most known problems. However, redundancy of word and sub-word occurrences in the document can be used to improve the recognition results. In this paper, we propose a highly accurate recognition system for printed old books. We use the combination of sub-word clustering and a LSTM neural network as a character recognizer to reduce the error rate. Due to the lack of information about the font faces, we manually label some part of the books. We show that the recognition error rate can be reduced noticeably by combining the results of the LSTM recognizer and sub-word clustering and rejecting the cases where the two methods differ in their prediction.

In this paper, a sparse method is proposed to synthesize the note-specific atoms for musical notes of different instruments, and is applied to separate the sounds of two instruments coexisting in a monaural mixture. The main idea is to explore the inherent time structures of the musical notes by a novel adaptive method. These structures are used to synthesize some time-domain functions called note-specific atoms. The note-specific atoms of different instruments are integrated in a global dictionary. In this dictionary, there is only one note-specific atom for each note of any instrument, resulting in a sparse space for each instrument. The signal separation is done by mapping the mixture signal to the global dictionary. The signal related to each instrument is estimated by a summation of the mapped note-specific atoms tagged for that instrument. Experimental results demonstrate that the proposed method improves the quality of signal separation compared to a recently proposed method