Kernel function is a normal method for image categorization to map high-dimension features into low-dimension spaces. Most state-of-art researches integrated kernels into Support Vector Machine (SVM) classifiers to solve classification problems. A novel Real Adaboost framework is proposed to involve kernel method to deal with classification. Hierarchical features PHOG and PHOW are used to describe shape and appearance information in multiple image scales first. Kernel function is then employed for evaluating features’ distance and constructing linear learner pool in kernel space. Real Adaboost is finally used to linear learners to obtain final image classifier. Experimental results show that our method significantly improves the image categorization performance.