A Method for Improving the Recognition Accuracy of Pattern Classification

We propose a simple and intuitive method for improving recognition accuracy in pattern classification. This model improves the average recognition accuracy by adjusting the number of pattern classification classes according to the characteristics of the pattern without being dependent on the algorithm. In this respect, the method proposed in this paper can be positively utilized in various applications. We verified that the overall average recognition accuracy can be improved by retraining after increasing the number of classes. Specifically, we retrained the data by assigning a new class to samples that are incorrectly recognized as a separated class, using handwritten digit recognition and text classification models. Our method shows some improvements in test accuracy by adjusting one class. In the text classification, we trained 4-class model based on our proposed idea with IMDb text corpus through detailed analyses. Our model recorded 99.70% as the average test accuracy, outperforming the state-of-the-art model which achieved a test accuracy of 96.21% so far.