Researches On Image Digital Watermarking Algorithm Based On Autoencoder

Image digital watermarking may be used for a wide range of applications,such as copyright protection,anti-counterfeiting and traceability.It is an important technology in the modern society with developed communication and is closely related to people’s daily life.Invisibility,robustness and capacity are the key and difficult points of current digital watermarking research.With the vigorous development of deep learning technology,more and more digital image processing methods,including digital watermarking have introduced deep neural networks.The digital watermarking algorithm based on deep neural network has achieved quite good performance in recent years,but there are still some shortcomings such as low capacity and poor robustness.Therefore,based on the existing digital watermarking algorithm and combined with the theoretical knowledge of deep learning and source channel coding,it’s proposed a block-wise image digital watermarking algorithm based on convolutional auto-encoder,which realizes the embedding and extraction of several binary images.The algorithm can be roughly divided into two modules: watermark embedding-extraction and source-channel joint coding.The watermark embedding-extraction module contains an end-to-end auto-encoder.It transforms the cover image into the DCT(Discrete Cosine Transform)domain,and embeds the watermark through the encoder while the decoder is responsible for the extraction of the watermark.Besides,there is a noise layer in the middle of the auto-encoder,which is used to improve the robustness of the algorithm.The source-channel joint coding module uses the variational auto-encoder based source-channel joint coding method NECST(Neural Error Correcting and Source Trimming)to compress the binary image into a smaller bit string,and decoded the noised bit string after it got through the embedding-extraction module.An attention based extraction algorithm is also proposed in this thesis.Inspired by the attention mechanism,this thesis adds an attention scorer to calculate the confidence of containing watermark at each image block in input images,and then focuses on the key areas with high confidence when extracting the watermark to exclude irrelevant information,improving the ability to resist clipping attacks.Finally,experiments are conducted to verify the effectiveness of the above algorithm.Compared with the current state-of-the-art algorithm,the algorithm in this thesis which can embed four 28×28 binary images in a 480×480 cover images,has better capacity,and the robustness to most watermarking attacks is slightly improved.