Aiming at insufficient robustness of depth map in multi-view 3D video with depth format, a 3D robust video watermarking algorithm based on depth map was proposed. Firstly, a depth map was divided into 4×4 nonoverlapped blocks, the mean square error of each block of pixels was calculated, and a threshold was set to distinguish the texture block and the flat block. Secondly, the block energy value of the texture block was calculated, and a threshold was set according to the calculated result to selectively embed the watermark bits. Finally, the transformed and quantized DC coefficients of each block were obtained and used to construct a 3×3 invertible matrix, then QR decomposition was performed on the invertible matrix and the watermark was embedded in the decomposed Q matrix. The proposed algorithm guarantees that the average Peak Signal to Noise Ratio (PSNR) is a constant, and the average Bit Error Rate (BER) under re-encoding attack with different Quantization Parameter (QP) values (25, 30, 35, 40) is 14.9%. Experimental results show that the algorithm has good robustness and embedding capacity, and has little impact on the quality of video.

Considering the low robustness of existing watermarking algorithms based on High Efficiency Video Coding (HEVC) standard, a robust video watermarking algorithm for HEVC based on intra-frame prediction modes of multi-partitioning was proposed. Firstly, in order to eliminate the intra-frame error propagation after embedding the watermark, embeddable regions were selected and the texture direction was calculated for 4×4 luminance blocks. Secondly, a scheme was proposed that the 33 angular prediction modes were divided into four pattern sets, which were recorded as follows:upper horizontal, lower horizontal, upper vertical, lower vertical. Finally, the four pattern sets were mapped to the values of the current and next to be embedded watermarks. Once the current 4×4 luminance blocks met the pattern sets, the current 33 angular prediction modes were truncated into the current pattern sets and the watermark was embedded. The watermark was extracted by the texture direction and four prediction pattern sets at decoding side. The experimental results show that, the average Peak Signal to Noise Ratio (PSNR) of the proposed algorithm is almost unchanged. In addition, the proposed algorithm achieves the Bit Error Rate (BER) of 14.1% under re-encoded attacks. Therefore, the proposed algorithm has low video distortion and can well resist the re-encoded attacks in robustness.

Considering low robustness of existing video watermarking algorithms for High Efficiency Video Coding (HEVC), a robust video watermarking algorithm based on texture direction was proposed in this paper. Depending on the value of watermarking, the intra-frame angular prediction modes were divided into horizontal modes and vertical modes, and the texture direction for each prediction unit was calculated when the splitting mode was N× N during compressing coding. Once the texture direction was consistent with the direction represented by watermarking, the 33 angular prediction modes of current prediction unit would be truncated to the horizontal or vertical direction prediction modes, and a best prediction mode which could decide whether watermarking was embedded would be decided on the basis of rate distortion cost function. The location of embedded watermarking would be recorded as a key for extraction at decoding side. The experimental results show that the proposed algorithm has low bitrate increasing and low video distortion, and Error Bit Rate (BER) still remains at low value after the attack of noise, filtering and re-encoding, which means the proposed algorithm can be used to protect video copyright.

The Scale Invariant Feature Transform (SIFT) algorithm is robust to the feature extraction of face image. However, the feature data derived by SIFT is of high dimension, and is difficult to be handled. Therefore, a Direct Locality Preserving Projections-SIFT (DLPP-SIFT) algorithm was proposed. In the algorithm, SIFT was used to extract feature, and the subspace method with Locality Preserving Projections (LPP) was utilized for dimension reduction. This algorithm solved locality preserving problem via simultaneous diagonalization; therefore, the singularity of the matrix was avoided. The experiments on ORL and FERET face databases show that the proposed algorithm reduces the computation complexity and matching time of features successfully, and is more robust than SIFT, Principal Component Analysis (PCA)-SIFT and LPP-SIFT methods.