Abstract: The environment of artificial tile classification is quite hostile, and the price of foreign industry vision equipment is expensive. As a result, it is important to develop indigenous classification systems and among which, the performance of classification algorithm plays a crucial role in determining the classification ability of the whole system. Traditional classification algorithms have limits in unlabeled instances estimate confidence and the classifier interference is large. To solve such problem, we proposed an algorithm based on constrained semi-supervised classification algorithms – Tri-training: search the instances which meeting the constraints of large number of unlabeled instances, increase the datasets of labeled instances, generate two strong classifiers, compose the integrated classifier as the final classifier for samples classification. Through experiments on real data sets, compared with the traditional algorithm, unlabeled instances estimate confidence increased 3%, classification accuracy increased 1.8%-3.3%.