A Watermarking Scheme Based on SVM and Tolerable Position Map

This paper presents a novel digital watermarking technique based on support vector machines (SVMs) and tolerable position map (TPM). The purpose of SVMs is two folds in this study. One is using SVM to identify tolerable positions for watermark embedding on the host image, and the other is using SVM to embed and extract watermarks. By simulating common image attacks on the host image, pixels which are invincible or vulnerable are identified and used for positive or negative samples for training an SVM. Apply this SVM can create a TPM for the host image. To embed and extract watermarks, we use a known binary sequence to train an SVM such that this SVM can be applied for embedding and extracting the watermark. In the proposed scheme, to improve robustness of attacks and image imperceptibility, the watermark is embedded according to the TPM and by asymmetrically tuning blue channels of the central and neighbor pixels. To further reducing extraction errors, the embedded watermark bits are re-modified if necessary according to classifying result of the trained SVM. Our scheme uses only 128 bits in training both SVMs, thus it is time efficient. Experiments show that the proposed scheme provides high PSNR of a watermarked image, low extraction error rate, and extremely robust to common image attacks.