Statistical and neural network algorithms are used to separate mine targets from clutter in side scan sonar images. In these images, a typical target usually contains in excess of 100 pixels filled with salt and pepper noise. This translates into a problem of classifying in a complicated high dimensional space, which is very difficult if not impossible to solve. Therefore, a typical mine detection algorithm contains three stages preceding the classification algorithm: noise reduction, clutter rejection, and feature extraction. These pre-processing steps would reduce the dimension of the feature space by an order of magnitude. Side scan sonar images are known to be contaminated with noise and mine like clutter. The major challenge is to select and measure the features of the potential targets. This is frequently done by fractal and/or Fourier analysis. Recently, wavelet analysis has also been used successfully as a tool for feature extraction. However, there are few analytical rules to guide the selection of features. In this paper, we investigate a new integrated feature extraction and classification algorithm that first enhances a potential target using variational based algorithms, and then transforms the enhanced image into a set of wavelet channels. We use the multichannel information as inputs to a feed-forward neural network. This new classifier has the advantage of extracting not only the local features but also the background features through higher scale wavelet channels. Results are compared for different network designs.