Normalization-Cooperated Gradient Feature Extraction for Handwritten Character Recognition

Normalization is a particular important preprocessing operation, and has a large effect on the performance of character recognition. One of the purposes of normalization is to regulate the size, position, and shape of character images so as to reduce within-class shape variations. Among various methods of normalization, moment-based normalizations are known to greatly improve the performance of character recognition. However, conventional moment-based normalization methods are susceptible to the variations of stroke length and/or thickness. In order to alleviate this problem, we propose moment normalization methods that use the moments of character contours instead of character images themselves to estimate the transformation parameters. Our experiments show that the proposed methods are effective particularly for printed character recognition.

The performance of handwritten Chinese character recognition (HCCR) has been greatly improved by using deep learning methods in recent years. But few people pay attention to the influence of writing style on it. In this paper, we aim to improve the performance of HCCR further by weakening the influence of different writing styles. We propose a writing style adversarial network (WSAN) which includes three parts: feature extractor, character classifier and writer classifier. In the training process, we first preprocess raw image with feature extractor. Afterwards, the learned features are fed into both the character classifier and the writer classifier. We apply joint optimization on the top of these two classifiers. Specifically, we minimize the loss value of the character classifier to achieve character recognition function. At the same time, we maximize the loss value of the writer classifier to reduce the influence of writing style in HCCR. The experimental results on CASIA-HWDB1.1 prove that the proposed WSAN has a promoting effect on HCCR. And the experiments on the offline HCCR competition dataset of ICDAR-2013 also give competitive results compared with other methods.

Optical character recognition for printed Indian text is a classical problem because of round nature of these characters and presence of ldquomaatrasrdquo. Recognition accuracy of optical character recognition engines depends heavily upon correct representation of characters obtained by feature extraction method. Thus feature extraction is heart of recognition engines. We propose a novel spline approximation based method for feature extraction in recognition engines. The spline approximation is implemented for Indian characters and the results obtained are very encouraging as compared to existing feature extraction methods.

LDA transformation is one of the popular feature dimension reduction techniques for the feature extraction in most handwritten Chinese characters recognition systems. The integration of the feature extraction and LDA transformation can be viewed as a two-directional feature transformation procedure, one is the pixel-level feature transformation by the summing up or blurring, another is by the LDA matrix, and the transformation coefficients are set empirically in the former. In this paper, we proposed a feature optimization method based on the gradient feature extraction by using the two-directional 2DLDA, which can find the optimal transformation coefficients in two directions. A series of experiments on the randomly selected 15 groups of the similar Chinese character samples from HCL2000 have indicated that, our method can effectively improve the recognition performance, the error rate reduction reaches 45.02% comparing to the traditional method, showing the effectiveness of the proposed approach.

This paper presents a novel geometrical scale and rotation independent feature extraction (FE) technique for multi-lingual character recognition (CR). The performance of any CR techniques mainly depends on the robustness of the proposed FE methods. Currently, there are very few scale and rotation independent FE techniques present in the literature which successfully extract the robust features from characters with noise such as distortion and breaks in the characters. Many FE methods from the literature failed to distinguish the characters which look similar in their appearance. So, in this paper, we have proposed a novel scale and rotation independent geometrical shape FE technique which successfully recognized distorted, broken, and similarly looking characters. Aside from the proposed FE technique, we've used crossing count (CC) features. Finally, we have combined the proposed features with CC features to make as Feature Vector (FV) of the character to be recognized. The proposed CR technique is evaluated using publicly available media-lab license plate (LP), ISI_Bengali, and Chars74K benchmark data sets and achieved encouraging results. To further assess the performance of the proposed FE method, we've used a proprietary data set containing nearly 168000 multi-lingual characters from English, Devanagari, and Marathi scripts and achieved encouraging results. We have observed better classification rates for the proposed FE method using publicly available benchmark data sets as compared to few of the CR FE methods from the literature.

Face recognition system, as any recognition process, depends highly on features extracted from face images. The selected features play a great role in deciding the recognition rate result. In this paper, a two-phase feature extraction and selection process is used for face recognition system. The process depends on histogram of Oriented Gradients (HOG) feature extraction and window size use to determine similarity between classes. Low number of features are used (big window size) to divide classes into small closed-similarity groups as first recognition phase. Then, the best matched class is found using larger number of features where differences between classes are bigger. The proposed method was applied to Essex face dataset using support vector machine (SVM) and Naïve Bayesian (NB) methods for comparison. The proposed method achieved 5% and 10% better recognition rate compared to SVM and NB respectively.

In handwritten Chinese character recognition, the performance of a system is largely dependent on the character normalization method. In this paper, a visual word density-based nonlinear normalization method is proposed for handwritten Chinese character recognition. The underlying rationality is that the density for each image pixel should be determined by the visual word around this pixel. Visual vocabulary is used for mapping from a visual word to a density value. The mapping vocabulary is learned to maximize the ratio of the between-class variation and the within-class variation. Feature extraction is involved in the optimization stage, hence the proposed normalization method is beneficial for the following feature extraction. Furthermore, the proposed method can be applied to some other image classification problems in which scene character recognition is tried in this paper. Experimental results on one constrained handwriting database (CASIA) and one unconstrained handwriting database (CASIA-HWDB1.1) demonstrate that the proposed method outperforms the start-of-the-art methods. Experiments on scene character databases chars74k and ICDAR03-CH show that the proposed method is promising for some image classification problems.

Degraded character recognition is a challenging problem in the field of Optical Character Recognition (OCR). The performance of an optical character recognition depends upon printed quality of the input documents. Many OCRs have been designed which correctly identifies the fine printed documents. But, very few reported work has been found on the recognition of the degraded documents. The efficiency of the OCRs system decreases if the input image is degraded. In this paper, a novel approach based on gradient pattern for recognizing degraded printed character is proposed. The approach makes use of gradient pattern of an individual character for recognition. Experiments were conducted on character image that is either digitally written or a degraded character extracted from historical documents and the results are found to be satisfactory.

The neocognitron is a hierarchical, multi-layered neural network capable of robust visual pattern recognition. The neocognitron acquires the ability to recognize visual patterns through learning. The winner-kill-loser is a recently introduced competitive learning rule that has been shown to improve the neocognitron’s performance in character recognition. This paper proposes an improved winner-kill-loser rule, in which we use a triple threshold, instead of the dual threshold used as part of the conventional winner-kill-loser. It is shown theoretically, and also by computer simulation, that the use of a triple threshold makes the learning process more stable. In particular, a high recognition rate can be obtained with a smaller network.

Machine learning offers promising solutions for the recognition of handwritten Javanese Nglegena script, which is crucial for preserving Indonesia's cultural heritage. This study explores the application of several supervised learning algorithms-K-Nearest Neighbors (KNN), Naïve Bayes, Decision Tree, and Random Forest-for classifying handwritten images of Nglegena Javanese script. Feature extraction is performed using a zoning technique, where each character image is divided into multiple zones (16, 25, 36, and 64) to capture local details. The extracted features are further processed using normalization methods, including Min-Max, Z-Score, and Binary normalization, to ensure uniform data distribution. The dataset, consisting of 600 images representing Javanese Nglegena characters, is split into training and testing sets using various ratios. Experimental results show that the combination of Naïve Bayes classification, 36-zone feature extraction, and Min-Max or Z-Score normalization achieves the highest accuracy of 65%. These findings demonstrate that optimizing zoning and normalization can significantly enhance the accuracy of machine learning models for Javanese script recognition. The research contributes to developing Optical Character Recognition (OCR) technology for Javanese script, supporting the digital preservation of Indonesia's historical and cultural heritage.

Ahstract-This study provides a novel brain-computer interface (BCI) approach for character recognition. The character recognition task is a two class classification problem. The key objective of the character recognition is to detect the P300 signal from the set of electroencephalogram (EEG) signals. The character is predicted from the detected P300 signal and row/column information of the oddball paradigm. The signal-to-noise ratio (SNR) of the electroencephalogram (EEG) signal is low. Ensemble of classifier is used to reduce the classifier variabilIty and enhance the SNR of the acquired signal. Here ensemble of SVMs (ESVMs) is used as classifier. The distribution of dataset is imbalanced due to its paradigm. A novel approach is applied in this work to balance the dataset. The proposed algorithm is evaluated on dataset IIb of BCI Competition II and dataset II of the BCI Competition III. The experimental results show that the proposed method outperforms the state-of-the-art character recognition performance.

The quality of the document images is a crucial factor for the performance of an Optical Character Recognition (OCR) model. Various issues from the input data hinder the recognition success such as heterogeneous layouts, skewness and proportional fonts. This paper investigated several algorithms for data pre-processing including image deskewing, table and document layout analysis to improve the accuracy of the OCR model and then built an end-to-end scanned document management system. We verified the algorithms using a well-known OCR software namely Tesseract. The experiments on a real dataset shown that our methods can accurately process document images with arbitrary angles of rotation, and different layouts. As a result, the accuracy by words of Tesseract can boost 23% for documents with complex structures. The quality of the output text allows to build a system to store and search documents efficiently. Index

The deep learning networks has become an active area in the realm of machine learning in recent years and some relevant models such as convolutional neural network (CNN) models have a good performance in handwritten character recognition. The MNIST is a typical open source data set, which only has ten kinds of handwritten digital characters from 0 to 9. Furthermore, the majority of the open source data sets in the current literature exclude arithmetic operators of "+" , "-" , "×" , "÷" , "=". In order to address this issue, firstly, we established a data sets of 22,200 characters, which includes fifteen categories. Specially, for each category, 1,200 characters is used for training and 280 characters is used for testing respectively. Secondly, we proposed a deep learning networks algorithm for characters recognition based on CNN. Thirdly, as an application, we further implemented four arithmetic operation using the recognized characters. The experimental results demonstrate that the proposed deep learning networks algorithm has shown strong ability for feature extraction and classification with the average accuracy of 97.42 %. More meaningfully, the four arithmetic operation in the present work has strong robustness and the potential applied value.

Communication is the ultimate of man’s search for conveying his ideas, emotions, and concepts. Dance is one of the media of communication through which dancers share notion of feelings, with the spectators through gestures, i.e., mudra. Gesture recognition propagates a concept without verbal speech or listening, and in dance recognition, the notion is transferred through various dance poses and actions. This activity in a way really paves way to enhance Indian Sign Language. This study focuses to solve the mudra resemblance in Bharatanatyam through a new system developed with image processing and classification technique using histogram of oriented gradient (HOG) feature extraction techniques and support vector machine (SVM) classifier. SVM classifies the features of HOG into mudras as text labels. Popular feature vectors such as scale-invariant feature transform (SIFT), speed up robust feature (SURF), and local binary pattern (LBP) are hardened against HOG for accuracy and speediness, and this innovative proposed concept is useful for online dance learners.

Cannabis is the second most common psychoactive drug in the world, and Thailand is the only country in Southeast Asia that allows people to use it. To maintain the integrity of different cannabis varieties and get the most out of their crops, growers, seed sellers, and farmers need to be able to accurately classify cannabis seed kinds. This paper presents a method for categorizing Thai cannabis seeds through the integration of Histogram of Oriented Gradients (HOG) feature extraction and Support Vector Machine (SVM) optimization, utilizing k-fold cross-validation and grid search methodologies. The suggested method worked well for smartly sorting different types of cannabis seeds. The regular SVM classifier got 94.11% accuracy, the k-fold cross-validation (K=10) got 94.00%, and the grid search optimization got 93.91%. These results indicate that the proposed method is both reliable and efficient for distinguishing cannabis seed varieties. Beyond its direct application to Thailand’s cannabis industry, the approach demonstrates the potential of combining HOG-based feature extraction with SVM optimization for other seed classification tasks in agriculture. By providing a scalable and accurate tool for seed identification, this work supports quality control, traceability, and productivity improvement in legal cannabis cultivation and trade.