Conventional blind watermark (WM) decoding schemes use correlation-based decision rules because of their simplicity. Drawback of the correlator decoders is their performance relies on the decision threshold. Existence of an undesirable correlation between the WM data embedded through a secret key and the host signal makes the decision threshold specification harder, especially in noisy channels. To overcome this drawback, we propose a SVM-based decoding scheme which is capable of learning the embedded WM data in wavelet domain. It is shown that both decoding and detection performance of the introduced WM extraction technique outperforms state-of-the-art correlation-based schemes. Test results demonstrate that learning in the wavelet domain improves robustness to attacks while reducing complexity.