Digital watermarking is a recent technology evolved to prevent illegal copy or reproduction of digital content. Most of the techniques developed use spatial and frequency domain for encoding the watermarks. These techniques fulfill the watermarking characteristics to varying degrees. There is a trade off observed between the information content and the fidelity of the cover image in almost all the works to a varying degree. This paper discusses a special scheme based on backpropagation neural network, which depends on small cover image parts to serve as inputs to a Backpropagation network and train them to produce corresponding small watermark image fragments. After training, the trained network weights are supplied with the cover image for the extraction of watermark. Small fragments of the cover image are taken to produce small fragments of the watermark image using trained weight matrix in the watermark extraction stage, which may be united to produce the original watermark image again. The watermark image is resistant to various image processing operations enhancing robustness of watermarking as the weights of the neural network remain unaffected by these operations.

Digital watermark, Neural Net, Backpropagation network