Training images are a bunch of images for which the required outcome is known. You input them to a program that analyzers their features and passes the features through a classification routine that determines the appropriate weights to use on the features in order to best achieve the required result.
Once you have used the training images to determine the best weights and procedures to follow, you then pass through a set of "test" images, for which you also know the outcome. The images are extracted and the procedures and weights determined above are applied to make a prediction of how the image should classify. Then the predicted classification is compared to the actual known classification to determine how well the procedures and weights do on analyzing data that was not originally used to determine the weights. If you got a high score on the training images but a low score on the test images then the implication is that the routine "overtrained" and has become too specific to the training images and not able to predict for images whose results are not known.
Once you are getting good scores on the training images and good scores on the test images as well, you can start using the procedures and weights to analyze data for which the outcome is not known.
