In recent years, we treat a lot of images easily because of the spread of smart phones, internet and digital cameras, etc. To search or put tags to images automatically, an object classification method with high accuracy is required and many object classification methods have been proposed. Recent some papers demonstrated the effectiveness of features obtained by the CNN trained from 1000 categories in the ImageNet. To improve the accuracy of features obtained by the pretrained CNN, we propose to use the combination of elements selected randomly from a feature vector obtained by the pretrained CNN. We do not know which combination is effective for a classification task. Thus, we compute many combinations of two or three randomly selected elements and trained a linear SVM with the combination feature. Only the combinations with large absolute weight of SVM are selected and used for final classification. In the experiments using the CIFAR10 dataset, we confirmed that accuracy is improved by the combination of randomly selected elements.