Recognition of low resolution text using deep learning approach

本論文關注的是電腦視覺中一個已充分研究過的議題，即光學文字識別。然而，我 們主要著重在一種非常特別的圖片類型:解析度非常低並且有大量失真與干擾的印刷中 文字。雖然使用卷積神經網路已能成功穩定識別高解析度印刷文字或手寫文字，然而， 對於品質非常低的印刷中文字仍有幾個挑戰，需要進一步分析研究。具體來說，我們的 資料集是點陣印刷機產生的 31,570 張文字圖片，包含模糊文字、缺少筆劃的文字以及 文字與其他文字或圖形重疊的文字圖片。為了有效地解決這些困難，我們實驗不同的深 層神經網路架構以及超參數，最後獲得辨識成果最佳的設置。在 1,530 類，平均解析度 為 16x18 像素的圖片中，top-1 和 top-5 的準確率分別為 71% 和 87%。Recent advances in deep neural networks have changed the landscape of computer vision and pattern recognition research significantly. Convolutional neural networks (CNN), for example, have demonstrated outstanding capabilities in image classification, in many cases exceeding human performance. Many tasks that did not get satisfactory results using conventional machine learning approaches are now being actively re-examined using deep learning techniques. This thesis is concerned with a well-investigated topic in computer vision, namely, optical character recognition (OCR). Our main focus, however, is a very specific class of input: printed Chinese texts with very low resolution and a significant amount of distortion/interference. Whereas the recognition of high-resolution texts, either printed or handwritten, has been successfully tackled using convolutional neural networks, the analysis of very low-quality printed Chinese texts poses several challenges that require further study. Specifically, our dataset consists of~31570~text images generated with dot-matrix printers, blurred texts, texts with missing strokes, and texts overlapping with other texts or graphs.To effectively address these difficulties, we have experimented with different deep neural networks with various combinations of network architectures and hyperparameters. The results are reported and discussed in order to obtain an optimal setting for the recognition task. The top-1 and top-5 accuracies are 71% and 87%, respectively, for input images with an average resolution of 16x18 pixels belonging to 1530 classes