Overflow problem is essential to reversible watermarking algorithms, as it would lead to great distortion or irreversibility if not properly handled. In general, the existing algorithms handle overflow problem with methods that take a lot changes to pixels, which consequently brings a possibly poor quality or a high complexity. In order to solve the overflow problem specifically with a higher quality and a lower complexity, and improve the security of watermark, we put forward a new reversible watermarking scheme. We take the strategy of twice embedding-first, the proposed scheme uses wavelet histogram shifting to embed the watermark information; second, a proposed low-distortion overflow processing algorithm (LDOPA) is implemented to process the overflow problem occurred during the first step, which iteratively scans the pixels, modifies the overflow pixels one by one, and embeds the modification record with its authentication information on the image. In addition, we apply logistic mapping, torus mapping, and CRC to improve the security of the watermark, allowing users to extract watermark only when they hold the correct password. The experimental results prove that the scheme realizes the complete reversibility of watermark and carrier image. Especially, the proposed LDOPA can be easily combined with the existing watermarking methods to further improve their performance. INDEX TERMS Reversible watermark, overflow handling, logistic mapping, torus mapping.

Mathematical expression retrieval is a key means of searching scientific contents in web and digital libraries. Because mathematical expressions have many different attributes compared with ordinary text, it is necessary to study the special retrieval methods including the indexing and matching model of mathematical expressions. In this paper, on the basis of the introduction of the existing math searching methods and the FDS based index, a mathematical expression matching model was proposed which realized the exact matching of formulas with three query modes called global query mode, local query mode and operational query mode. The algorithms of the query modes were given respectively. A prototype system based on the proposed model was implemented and the comparison experiments were carried out. The experimental results show that the proposed matching model simultaneously realized matching formulas in exact mode and reducing time and space consumption of retrieving to an acceptable degree. It is effective for searching math content in relative digital mathematics library.

No abstract

Distantly supervised relation extraction (RE) automatically aligns unstructured text with relation instances in a knowledge base (KB). Due to the incompleteness of current KBs, sentences implying certain relations may be annotated as N/A instances, which causes the socalled false negative (FN) problem. Current RE methods usually overlook this problem, inducing improper biases in both training and testing procedures. To address this issue, we propose a two-stage approach. First, it finds out possible FN samples by heuristically leveraging the memory mechanism of deep neural networks. Then, it aligns those unlabeled data with the training data into a unified feature space by adversarial training to assign pseudo labels and further utilize the information contained in them. Experiments on two wildlyused benchmark datasets demonstrate the effectiveness of our approach.

Distantly supervised relation extraction (RE) automatically aligns unstructured text with relation instances in a knowledge base (KB). Due to the incompleteness of current KBs, sentences implying certain relations may be annotated as N/A instances, which causes the socalled false negative (FN) problem. Current RE methods usually overlook this problem, inducing improper biases in both training and testing procedures. To address this issue, we propose a two-stage approach. First, it finds out possible FN samples by heuristically leveraging the memory mechanism of deep neural networks. Then, it aligns those unlabeled data with the training data into a unified feature space by adversarial training to assign pseudo labels and further utilize the information contained in them. Experiments on two wildlyused benchmark datasets demonstrate the effectiveness of our approach.