Dermoscopic imaging is one of the most effective tools for improving the rate of early melanoma recognition. Our research focuses on three components of an automated melanoma recognition system, including lesion segmentation, representation, and classification. Our main contributions include the following: (1) We propose a new segmentation method combining graph partitioning active contours and narrow band energy. Our method provides highly competitive results compared to other segmentation algorithms on a set of over 200 images; (2) We employ mid-level features for lesion representation and evaluate various sets of feature combinations and selection schemes for these features. We also compare different sampling strategies for mid-level feature generation. In addition, we investigate the use of graph models to integrate spatial information with mid-level features. Experiments on a set of over 1,500 images show that spatial information can significantly improve melanoma recognition accuracy; (3) We formalize the problem of recognizing high-level anatomical features as a multi-instance learning model and compare it against traditional single instance learning. We further propose a newly developed learning model (with proven error bound) that can use these additional anatomical features provided by domain experts. Comparison against relevant learning models on a set of 360 lesion images shows that our method can effectively use the additional information.