Recognition of Japanese Connected Cursive Characters Using Multiple Softmax Outputs

Ueki, Keisuke; Kojima, Tomoka; Mutou, Ryou; Nezhad, Rostam Sayyed; Hagiwara, Yasuaki

doi:10.1109/mipr49039.2020.00033

Cited by 6 publications

(5 citation statements)

References 4 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…Many models have been designed and implemented to recognize Kuzushiji characters from Japanese historical documents. Some wellknown recognition models are RU-Net [40], AED [41], AACRN [45], 2DCbpn [48], MSOs [47], and RAED [13].…”

Section: Discussionmentioning

confidence: 99%

See 1 more Smart Citation

DKNet: Deep Kuzushiji Characters Recognition Network

Singh

Aravinda²,

Kaur

et al. 2022

IEEE Access

View full text Add to dashboard Cite

Kuzushiji, a cursive writing style, had been extensively utilized in Japan for over a thousand years starting from the 8 th century. In 1900, Kuzushiji was not included in regular school curricula due to the change in the Japanese writing system. Nowadays Japanese natives are unable to read historical books that were written using Kuzushiji language. Therefore, libraries and museums have decided to build digital copies of the documents and books that were written in Kuzushiji language. Due to a limited number of trained experts, researchers have utilized machine and deep learning models to convert historical documents and books into a modern script that can be easily read by human beings. However, the existing deep learning techniques suffer from over-fitting and gradient vanishing problems. To overcome these problems, an efficient deep Kuzushiji characters recognition network (DKNet) is proposed. Initially, to remove noise from the training images, a trilateral joint filter is applied. Contrast limited adaptive histogram equalization (CLAHE) is then applied to enhance the visibility of filtered images. Thereafter, a pre-trained MobileNet is utilized to extract the features of Kuzushiji characters. MobileNet's final layers are removed, including the fully connected layer and softmax. The flatten layer is then applied to the input. A fully connected classification layer is then used with Rectified linear units (ReLUs) and dropouts. Dropouts are used to generalize the model, thus preventing the over-fitting problem. Finally, the softmax activation function is employed to provide the recognition results. To test the proposed model, actual documents are first segmented by using the proposed Maximally stable extremal regions (MSERs) and convexhull-based segmentation approach. Segmented characters are then recognized using the trained DKNet. Extensive comparative analyses reveal that DKNet achieves better performance than the competitive models in terms of various performance metrics. An efficient Application Programming Interface (API) is also designed for Japanese Kuzushiji ancient heritage character recognition to help the end-users. INDEX TERMSMaximally stable extremal regions, Deep Learning Models, Kuzushiji character recognition, MobileNet, InceptionNet, ResNet I. INTRODUCTIONK UZUSHIJI, a cursive writing style, had been extensively utilized in Japan for over a thousand years starting from the 8 th century. But, in 1900, Kuzushiji was not included in regular school curricula due to the change in Japanese writing system. Nowadays Japanese natives are unable to read historical books that were written using Kuzushiji language. Therefore, libraries and museums have decided to build digital copies of the documents and books that were written in Kuzushiji language. But, due to a limited number of trained experts, researchers have utilized machine and deep learning models to convert historical documents and

show abstract

Section: Discussionmentioning

confidence: 99%

“…In [46], the Deep learning model(DLM) was applied to recognize the Kuzushiji characters. Ueki et al [47] proposed a recognition system for the Kuzushiji characters by utilizing Multiple softmax outputs (MSOs). In [48], a 2-dimensional Context box proposal network (2DCbpn) was proposed to identify the Kuzushiji Characters.…”

Section: Literature Surveymentioning

confidence: 99%

DKNet: Deep Kuzushiji Characters Recognition Network

Singh

Aravinda²,

Kaur

et al. 2022

IEEE Access

View full text Add to dashboard Cite

show abstract

“…CNN memecah gambar menjadi beberapa pixels. Setiap pixels hasil dari konvolusi kemudian dijadikan sebagai inputan untuk mendapatkan hasil representasi fitur [14]. Proses ini dijadikan sebagai langkah pengenalan obyek saat obyek tersebut muncul.…”

Section: Gambar 2 Convolutional Neural Networkunclassified

Deteksi Karakter Hiragana Menggunakan Metode Convolutional Neural Network

Rachmawanto

Andono

2022

j. nas. pendidik. teknik. inform.

View full text Add to dashboard Cite

Salah satu jenis huruf dasar yang digunakan dalam Bahasa Jepang ialah Hiragana. Dalam penulisan Hiragana memiliki aturan guratan dasar atau berbentuk garis – garis dan coretan melengkung (kyokusenteki), penulisan dari garis atas ke bawah atau dari kiri ke kanan. Aturan tersebut harus diikuti dan diperhatikan. Diusulkan penggunaan Metode Convolutional Neural Network (CNN) untuk melakukan pengenalan terhadap tulisan tangan karakter Hiragana. Pada tahap preprocessing, terdapat proses segmentasi yang menggunakan metode thresholding, setelah itu dilakukan proses menghilangkan noise, mengubah ukuran, dan proses normalisasi dengan cara memotong gambar dataset. Pada tahap pengujian, digunakan Metode Adam Optimizer sebagai alat untuk menguji metode yang digunakan dimana akan menghasilkan nilai akurasi. Dengan menggunakan 1000 dataset gambar yang terdiri dari 50 karakter, masing-masing karakter memiliki 20 sampel gambar. Dari 1000 data tersebut, telah dilakukan variasi data dengan melakukan split dataset. Dalam penelitian ini dilakukan 2 kali percobaan dengan variasi data 70:30 dan 60:40. Akurasi yang didapat yaitu 86,5% dan 83%.. Akurasi yang di dapat menggunakan split dataset 70:30 ternyata menghasilkan prosentase lebih tinggi di banding menggunakan split dataset 60:40.

show abstract

“…The authors reported that the recognition rate of a single character was approximately 92%; however, the recognition rate of three characters was only approximately 76%. Similarly, a method for recognizing a string of three consecutive characters using a sliding window and BLSTM was proposed [12]. The authors used the tendency in which the maximum output probability of a neural network is not particularly high for a misaligned character image but is high for an accurately aligned image, and increased the recognition rate to 86% by integrating multiple results.…”

Section: Representative Research On Kuzushiji Recognitionmentioning

confidence: 99%

“…In addition, the training data were augmented using techniques such as combining multiple images by applying mixup [45] and random image cropping and patching (RICAP) [46], and adding some noise to the image by random erasing. Because Furigana 12 is not a recognition target, post-processing such as the creation of a false positive predictor is used for its removal.…”

Section: B Kaggle Competitionmentioning

confidence: 99%

Survey on Deep Learning-based Kuzushiji Recognition

Ueki

Kojima

2020

Preprint

Self Cite

View full text Add to dashboard Cite

Owing to the overwhelming accuracy of the deep learning method demonstrated at the 2012 image classification competition, deep learning has been successfully applied to a variety of other tasks. The high-precision detection and recognition of Kuzushiji, a Japanese cursive script used for transcribing historical documents, has been made possible through the use of deep learning. In recent years, competitions on Kuzushiji recognition have been held, and many researchers have proposed various recognition methods. This study examines recent research trends, current problems, and future prospects in Kuzushiji recognition using deep learning.

show abstract

Recognition of Japanese Connected Cursive Characters Using Multiple Softmax Outputs

Cited by 6 publications

References 4 publications

DKNet: Deep Kuzushiji Characters Recognition Network

DKNet: Deep Kuzushiji Characters Recognition Network

Deteksi Karakter Hiragana Menggunakan Metode Convolutional Neural Network

Survey on Deep Learning-based Kuzushiji Recognition

Contact Info

Product

Resources

About