Key Character Detection Based on Deep Neural Network for Document Images

In this paper, a key character detection method was proposed based on deep neural network for document images. The proposed method is developed based on RetinaNet. The first step is the ROI detection. After obtaining the ROI, RetinaNet composed of convolutional layer and feature pyramid network is used to detect key characters. To train the model of RetinaNet, data augmentation is used to generate a large image dataset. To evaluate the proposed method, IoU (intersection over union) and mAP (Mean Average Precision) are used as performance measurements. For performance evaluation, some document images are selected for testing. Experimental results show the average precision rates of detecting key characters by using the proposed method are at least 78% and mAP is up to 91%. The experimental results demonstrate that the proposed method can effectively localize and recognize these key characters in document images.