Recently, a constrained Linear Discriminant Analysis (LDA) algorithm is introduced and gained popularity. However, this algorithm is not applicable in the environment with large amount of data points or when the data point arrive in a sequential manner. In this paper, we aim to propose an incremental version of this algorithm called Incremental Constrained Discriminant Component Analysis (ICDCA) to reduce the computational cost of this algorithm in large datasets.The ICDCA updates the within class scatter matrix and between class scatter matrix instead of calculating it from scratch. This change significantly reduces the computational cost of feature extraction process while keep the accuracy of such features as close as possible to offline version of this algorithm. In the end the effectiveness of ICDCA is compared to other recently proposed incremental LDA. To ensure the reliability of these experiments, they are repeated with several UC I data set. In these comparisons, advantage of ICDCA in the accuracy and speed is demonstrated.

Keywords