A novel watermarking scheme is proposed that could substantially improve current watermarking techniques. This scheme exploits the features of micro images of watermarks to build association rules and embeds the rules into a host image instead of the bit stream of the watermark, which is commonly used in digital watermarking. Next, similar micro images with the same rules are collected or even created from the host image to simulate an extracted watermark. This method, called the features classification forest, can achieve blind extraction and is adaptable to any watermarking scheme using a quantization-based mechanism. Furthermore, a larger size watermark can be accepted without an adverse effect on the imperceptibility of the host image. The experiments demonstrate the successful simulation of watermarks and the application to five different watermarking schemes. One of them is slightly adjusted from a reference to especially resist JPEG compression, and the others show native advantages to resist different image processing attacks.