基于改进残差学习的东巴象形文字识别

基于深度学习模型的东巴象形文字识别效果明显优于传统算法,但目前仍存在识别字数少、识别准确率低等问题.为此本文建立了包含1387个东巴象形文字、图片总量达到22万余张的数据集,大幅度增加了可识别字数,并辅助提高了东巴象形文字的识别准确率.同时,本文根据东巴象形文字相似度高、手写随意性大的特点,选择ResNet模型作为改进...

Dongba pictographs recognition based on improved residual learning

Dongba pictographs recognition based on deep learning model has better recognition effect than that of traditional algorithms. However, these methods have disadvantages such as small number of recognizable Dongba pictographs and low recognition accuracy. Aiming at these problems, in this study, we build a novel dataset of Dongba pictographs that contains 1387 Dongba pictographs and more than 220 thousand images. Therefore, the number of recognizable Dongba pictographs is greatly increased and the Dongba pictographs recognition accuracy is improved. Since Dongba pictographs are characterized by high similarity and random writing, ResNet is adopted as an improved network structure. Moreover, we design a residual shortcut connection and the number of convolutional layers and introduce the max-pooling into the ResNet to improve down-sampling. The experimental results demonstrate that the improved ResNet model can recognize more Dongba characters, and has achieved the highest recognition accuracy 98.65% in our dataset.