紀要論文

情景内カラー文字認識のための最適2値化と変形耐性画像マッチング

Recognition of Color Characters in Scene Images via Optimal Binarization and Distortion-tolerant Image Matching

jpn

24321192

https://doi.org/10.15002/00021943

法政大学大学院情報科学研究科

法政大学大学院紀要. 情報科学研究科編

14

1

6

2019-03-31

Version of Record

In this paper, we propose the method of recognition of color characters in scene images. In this method, we apply two key ideas to recognition. The first idea is optimal binarization, which separates the image into the character region and the background region. We use K-means clustering in Lab color space to generate multiple segmented images. After generation, these segmented images are classified to correctly binarized character image or not, by the convolutional neural network. All of the segmented images are classified, one of binarized image is selected as the optimal binarized character image among them. The second idea is distortion-tolerant image matching. Distortion indicates rotation, expansion, shearing, and translation, expressed by an affine transformation. In this matching, the correlation value is calculated between the target image and the template images first. Then, the optimal affine transformation is applied to the template images, to make a higher correlation value. These steps are repeated while the maximum correlation values are gained. Finally, we select k samples which gained higher correlation values and classify the target image using k-NN classification method. Experimental results made on the public color character image dataset “the Chars74K” show that the proposed method achieves 96.6% of correctly binarized character image selection rates and 73.9% of character recognition rates.

Article

AA12746425