Chinese document image retrieval based on recognition candidates

For the sake of the low recognition rate for degraded Chinese document, the retrieval performance is not good if directly based on OCR result. In this paper, an indexing method with n-gram and recognition candidates is proposed to improve the performance of retrieval. For ease of test, this paper also presents a method to automatically generate ground-truth of imaged document, synthesized degraded document image and ground-truth of recognition candidates. Several synthesized document image collections on large-scale are built and used, and the experimental results show that the retrieval performance are improved for both collections with high or low OCR error rates.