High-performance reversible data hiding based on ridge regression prediction algorithm.

• The ridge regression predictor is utilized to minimize the sum of squared residuals between the prediction pixels and the target pixels for reducing prediction error. • The ridge regression predictor only decreases the weights of abnormal samples and weakens the influence of abnormal samples on prediction, avoiding the overfitting phenomenon. • the proposed method also selects the nearest neighbouring pixel around the target pixels as the training sets and the supported sets of ridge regression predictor, which further improves the prediction accuracy. An effective error prediction algorithm is the key to improving the embedding performance of reversible data hiding schemes. In this paper, high-performance ridge regression predictor-based reversible data hiding (RDH) is proposed. The ridge regression predictor is an adaptive predictor that adds L2 regularisation to minimise the residual sum of squares between the prediction pixels and the target pixels. The L2 norm as a penalty function decreases the weights for the prediction coefficients of unimportant pixels. In other words, the ridge regression predictor limits prediction coefficients that have negative or no influence on predicting the target pixels (abnormal samples). The ridge regression predictor allows the prediction coefficients to be small, which avoids the overfitting problem and enhances tamper-resistance and generalisation ability. In addition, to increase the prediction accuracy of the ridge regression predictor, the proposed method employs small samples to obtain more accurate prediction values. The neighbouring pixels closest to the target pixels are selected as the training sets and supported sets during the prediction process. In summary, the ridge regression predictor can generate an error plane that is more suitable for embedding, thereby improving the embedding performance of RDH. Extensive experimental results also show that the proposed method is superior to the state-of-the-art RDH schemes in terms of prediction accuracy and embedding performance. [ABSTRACT FROM AUTHOR]