Abstract-Incremental learning is a machine learning paradigm that the training dataset grows dynamically at runtime and the prediction accuracy can be improved incrementally. Since reallife datasets are usually open-ended and dynamic, incremental learning is important to many image classification applications. In existing incremental algorithms, KNN (K Nearest Neighbors) has the advantage of dealing with the hugely and incrementally multiclass nature of visual object recognition. However, it provides good prediction only when the query image is similar to one of the images in the dataset, which results in low accuracy in many cases. In this paper, we propose a hybrid incremental learning algorithm named IKNN-SVM which can improve the recognition accuracy while reserving the advantages of KNN, i.e., quickly learning from the increasing number of data. To achieve this goal, our incremental algorithm combines SVM (Support Vector Machine) with high recognition accuracy and KNN together. And to promote the incremental learning efficiency, IKNN-SVM introduce kd-tree to implement fast searching. The experimental result based on the open dataset Caltech256 shows that the recognition accuracy of our method improves by 12%-26% compared with incremental KNN but with short time consumption.