A Statistical Approach for Photo Quality Assessment

Abstract

This study proposes a photo quality assessment method based on the spatial relations of image patches. High aesthetic quality images and low aesthetic quality images are used to explore elements both of beauty and of less attractive features. To explore the components of high-quality photos, the image is decomposed into patches based on color information. Then the low-level features, color moment, and histogram of oriented gradients (HOG) are extracted for each patch. Because the aesthetic features differ depending on the type of scene depicted, each photo is assigned to one subtopic before implementing the learning model. Distinguished from prior research, which modeled the spatial relations of image patches from high-quality photos only, our model also generates negative models from low-quality photos to provide more discriminate assessment results. Spatial information about the location and size of image patches is modeled by the Gaussian mixture model (GMM), and the likelihood probabilities from both the positive and negative context models are integrated into one assessment score. The experimental results demonstrate that adding low-quality photos can provide significant improvement for photo quality assessment.