Here is the dataset and the correct classes for each image. This is the only data you will get, so use it wisely. A good idea is to split it up in two sets, one training set and one validation set. The first one should be larger and is used to train your classifier and the later to measure the performance of the classifier so you know when to stop training.
The noise in the images is from a normal distribution with zero mean. The noise in each pixel is independent of the noise in all other pixels.
There is no formal description of the fileformat for the trainingdata, but it's briefly described here. The test-set used will have the same size and color range. The only thing that will differ from the training set is the number of images and the image data.
Read answers to some questions from last year here.