Wavelet networks have emerged as a powerful tool for nonparametric estimation. It is a method implementing inverse discrete wavelet transform with coefficient optimization techniques from machine learning field. However, conventional ways to construct wavelet networks are based on empirical risk minimization (ERM) principle, which has been proven not as robust as structural risk minimization (SRM) principle. Thus, to explore the optimal architecture of wavelet networks, we constructed wavelet networks based on SRM principle. This paper describes the kernel-based way to optimize the architecture of wavelet networks. Based on the frame theory, wavelet kernel functions are found. After that, the wavelet network is constructed with support vectors generated by the wavelet kernel functions.