Co-forest is a classical semi-supervised collaborative training algorithm. Aming at solving the unavoidable mislabeling problem and the defect of decision tree classifier with low performance. In this paper, a novel Improved Co-forest (Impr-Co-Forest) algorithm is proposed to address the above issues. First, considering the creation and selection of optimized decision trees, in training process, we applied a weight scheme into adaboost method in order to pay more attention to the samples that are difficult to classify. Because newly labeled data are not always valid, the Out-Of-Bag-Error (OOBE) of single decision tree is compared with the decision trees with smaller OOBE. Then, to solve the mislabeling problem throughout the colabeling iterations, a dual-confidence estimation method is proposed for the newly labeled data that is effectively chosen to update classifiers. Finally, the weighted vote, which replaces the simple majority vote, achieves the estimation of probability for each class. Experimental results on eight UCI datasets and Pascal VOC show that the proposed Impr-Co-Forest algorithm has better classification performance than both supervised and semi-supervised algorithms.