Lagrangian support vector regression based image watermarking in wavelet domain

Abstract

To enhance the imperceptibility and robustness against image processing operations, the advantage of artificial neural network (ANN) and machine learning algorithms such as support vector regression (SVR), extreme learning machine (ELM) etc. are employed into watermarking applications. In this paper, Lagrangian support vector regression (LSVR) based blind image watermarking scheme in wavelet domain is proposed. The good learning capability, high generalization property against noisy datasets and less computational cost of LSVR compared to traditional SVR and ANN based algorithms makes the proposed scheme more imperceptible and robustness. Firstly, four sub images of host image are obtained using sub sampling. Each sub image is decomposed using discrete wavelet transform (DWT) to obtain the low frequency subband. Low frequency coefficients of each sub image are used to form the dataset act as input to LSVR. The output obtained by trained LSVR is used to embed the binary watermark. The security of the watermark is enhanced by applying Arnold transformation. Experimental results show the imperceptibility and robustness of the proposed scheme against several image processing attacks. The visual quality of watermarked image is quantified by the peak-signal-to noise ratio (PSNR) and the similarity between the original and extracted watermark is evaluated using bit error rate (BER). Performance of the proposed scheme is verified by comparing with the state-of-art techniques. Keywords-Wavelet transform; Lagrangian support vector regression; Peak signal to noise ratio; Bit error rate;I.

Blind Image Watermarking Based on Adaptive Data Spreading in n-Level DWT Subbands

“…Meanwhile, the values of our scheme were measured by applying our scheme to the same image set and the same set of attacks as the corresponding existing scheme. For the watermark data, we used the same size of data (marked as "-" in Table 4) as the one used in the scheme to be compared: 32×16 for [17], 32×32 for [9,16], and 64×64 for [14,15]. First we compare with [17], which is in Table 5.…”

“…Finally [9] is compared with ours in Table 9, whose robustness was assessed by BER. This scheme included some pixel-value change attacks and the 25% cropping geometric attack.…”

“…The watermarking data was embedded in the 2 -level HH subband. In [9,10], level-1 and level-2 2D DWT were also used. In [10] the resulting HH subband was used to embed a 64×64 watermark data.…”

“…In [10] the resulting HH subband was used to embed a 64×64 watermark data. In [9] the image was divided into 4 subimages, each of which was 1-level DWTed. A 32×32 watermark data was embedded into the resulting LL subbands by adding or subtracting 20 to embed one watermark bit.…”

“…A quantization index modulation (QIM) was used in [9,[13][14][15][16][17]. It was shown that QIM had high strength to some peculiar attacks but was weak to the geometric attacks.…”