Chinese character detection using modified Single Shot Multibox Detector
Junhwan Ryu, Sungho Kim · International Conference on Control, Automation and Systems · 2018
In this paper, we propose a deep learning based method for the detection of Chinese characters. Previous studies have been based on traditional methods such as text-line and character segmentation, text region detection, and local projection. These conventional methods are severely affected by the noise level of the image. The proposed method utilize a network with a modified Single-Shot Multibox Detector (SSD), which is the state-of-the-art network suitable for Chinese character dataset. We used Resnet instead of the existing VGG network in the base network. The aspect ratio and scale of the default box in SSD, which excludes the last layer and predicted, showed a maximum detection rate of 99% for the corresponding data set.