When SVM is adopted for image annotation, we know that high quality sample images will improve image recognition accuracy. Images with the same visual/semantic features are adopted as positive sample images, and images with different visual/semantic features are adopted as negative sample images. But it is labor intensive in high quality sample images selection, especially when collecting by visual features. Most researchers randomly choose positive and negative sample images for classifier training. In many applications, adopting different negative sample image datasets will vary annotation accuracy. In this research, we will discuss the accuracy between different negative sample images dataset collected by semantic features. We adopted Image Net as image dataset in this study, and we adopted Word Net for building semantic hierarchical tree. Semantic hierarchical structure tree is adopted to calculate the distance between each node. Then we adopt this distance relationship to prepare positive and negative sample images. We prepare one baseline method and suggest six different negative sample images selection methods for experiment. The binary SVM classifier training and prediction is implemented to compare the accuracy and Mean Reciprocal Rank (MRR) between baseline and each proposed method. Our results show that if we select uniform amount of negative sample images in each distance in the semantic hierarchical tree, we will achieve highest accuracy.