A statistic approach for photo quality assessment

Abstract

This study proposes a photo quality assessment based on the spatial relations of image patches. In order to investigate the components of high-quality photos, the image is decomposed into patches based on the color information. Then the color moment and histogram of oriented gradients (HOG) are extracted for the feature representation. Because the diverse types of photos, the photo with the segmented patches is assigned to a subtopic before further modeling. Different from the prior researches which model the spatial relations of image patches obtained from high quality photo, in our work the negative models are learned from the low quality photos as well to provide more discriminate assessment results. Note that the spatial information of location and size of image patch is modeled by Gaussian mixture model (GMM), and the likelihood probabilities in accordance with the positive and negative context models are integrated as the assessment score. The experimental results demonstrates that the usage of the low-quality photos can provide the significant improvement and the proposed system have the promising potential for the photo quality assessment.