Image databases for quality assessment are helpful to evaluate the performance of objective assessment methods. Recommendations in regard to the constitution of databases and experimental methods of the subjective assessment have been proposed to ensure the database a good ground truth for the validation of objective quality assessment methods. However, these restrictions make databases scale-limited by covering small number of scenes distorted by few levels. To enrich IQA databases and increase the generalization capability of IQA models, we devise an effective image augmentation method. The two-stages scheme consists of the image-label pairs generation by minimizing the free energy between the pristine image and its augmentation as well as the distortion level interpolation which is based on the monotonicity of the perceptual quality with the severity of distortion. The experimental results show the ability of the augmented database to improve the prediction accuracy of learning-based no-reference image quality assessment metrics which in turn demonstrates the effectiveness of our method.