We show naive Bayes models of the melanoma skin cancer represented by dermoscopy images from two different repositories. The dermoscopy images of each data set are recursively analyzed by the Mallat wavelet tree transform to extract a set of spatio-frequency filters, which are used to build energy based (rotation invariant) features. Such classifiers are built in different wavelet bases and (for one data set) in three image resolutions. Those simple models show varying classification performance, but some wavelet bases are preferable to differentiate between malicious (melanoma) and benign (displastic nevus) lesions and keep it in reduced image resolutions. The presented research contributes to the feature extraction (wrapper) methods.