Programme and Methodology

Our work focuses on three cases. Each employs photometric stereo which requires the use of at least three images taken using different illumination source positions. We assume that we can obtain these photometric image sets during training in all cases. However, only in the first case do we assume that we can obtain photometric data during the classification sessions.

Case 1 – Full photometric training and classification

This case assumes the use of photometric image sets both for training and classification sessions. The use of an ‘illumination rig’ (a frame containing computer controlled illumination sources at known positions) together with a single fixed camera means that this is a simple, reliable and fast method of obtaining data. It can be used in many industrial inspection applications.

From the photometric image sets we can obtain estimates of the partial derivatives p(x,y) and q(x,y) of the surface, together with estimates for the reflectance coefficient at each point (x,y) on the surface. (These are obtained both during training and classification sessions. )

Classification is performed on features derived directly from the gradient and reflectance coefficients. So, we do not compare, for example Gabor coefficients computed from intensity values I(x,y) associated with a single image, rather we compare features computed directly from p, q, and albedo estimates. In practice we statistically derive discriminants from the training sessions and use these to construct the classifier. In other words, our invariant features are computed from surfaces’ three-dimensional and reflectance characteristics. If rotation insensitive filters are used on gradient magnitudes then a true surface-rotation-invariant classifier results.

Case 2 – Photometric training with single image classification

Here we assume that photometric image sets are only available during training, but that we know the viewing geometry and illumination conditions under which the single test image was captured. Such a case may arise in a remote sensing application. As in case 1 above, we estimate the partial derivatives and reflectance coefficients for each of the surface classes during training.

Using the imaging and illumination geometry under which the test image was obtained we can synthesise one view for each of our surface classes. Conventional texture features can be derived from the synthesised images and used to train a classifier. The classifier then can be used to classify the test image.

This approach can deal with a wide range of illumination conditions but cannot deal with unknown rotations of the test surface.

Case 3 – Photometric training with single image classification (unknown imaging conditions)

This case is identical to case 2 except that we assume the illumination conditions used to obtain test images are not known. Here we must have a scheme where for each texture class, we identify a set of ‘representative views’. Each set of ‘representative views’ will be dependent upon the range of illumination conditions considered, and the number and type of texture classes involved. Discriminants must then be constructed for each representative view for each texture class. Classification sessions would then simultaneously identify both the texture class and the illumination conditions. This is the most ambitious of the cases and it is likely that both the range of illumination conditions and texture classes will have to be restricted.