Text content can have different visual presentation ways with roughly similar characters. While conventional text image retrieval depends on complex model of OCR-based text recognition and text similarity detection, this paper proposes a new learning-based approach to text image retrieval with the purpose of finding out the original or similar text through a query text image. Firstly, features of text images are extracted by the CNN network to obtain the deep visual representations. Then, the dimension of CNN features is reduced by PCA method to improve the efficiency of similarity detection. Based on that, an improved similarity metrics with article theme relevance filtering is proposed to improve the retrieval accuracy. In experimental procedure, we collect a group of academic papers both including English and Chinese as the text database, and cut them into pieces of text image. A text image with changed text content is used as the query image, experimental results show that the proposed approach has good ability to retrieve the original text content.