1. Segment the leaf from the background. Even though we require users to place the leaf on a white background, this is still a very tough problem because of lighting, shadows, viewpoint, blur, focus, etc. We've tried all published segmentation algorithms and nothing works. We wrote our own, using EM on the HSV colorspace as the basis for it (plus lots of heuristics for leaves). It works okay, but it definitely needs the most work.
2. Extract some features from the contour of the leaf. We use histograms of curvature over scale. The idea is that some leaves have the same rough shape, but differ at a fine scale (e.g., serrated vs. smooth leaves). Others are similar at a fine-scale but look quite different overall. To distinguish all of these cases, we estimate "curvature at a scale" at each point on the contour and build a histogram of curvature values. We then repeat this at coarser and coarser scales and concatenate all histograms together to get a large feature vector.
3. Recognition/matching: We use a simple nearest-neighbors classifier to match the feature vectors to our ground-truth database of labeled leaves. We return the top 20 ranked species or so.