Affective image classification has attracted much attention in recent years. However, the production of more exact classifiers depends on the quality of the sample database. In this study, we analyzed various existing databases used for affective image classification and we tried to improve the quality of the learning data by combining existing databases in several different ways. We found that existing image databases cannot cover the overall range of the arousal-valence plane. Thus, to obtain a wider distribution of emotion labels from images, we conducted a crowd-sourcing-based user study with Amazon Mechanical Turk. We aimed to construct several different versions of affective image classifiers by using different combinations of existing databases, instead of using one. We used low-level features in our classification experiments to explore the discriminatory properties of emotion categories. We report the results of intermediate comparisons using different combinations of databases to evaluate the performance of this approach.