Feature Construction Using Genetic Programming for Classification of Images by Aesthetic Value

Classification or rating of images according to their aesthetic quality has applications in areas such as image search, compression and photography. It requires the construction of features that are predictive of the aesthetic quality of an image. Constructing features manually for aesthetics prediction is challenging. We propose an approach to improve on manually designed features by constructing them using genetic programming and image processing operations implemented using OpenCV. We show that this approach can produce features that perform well. Classification accuracies of up to 81% on photographs and 92% on computationally generated images have been achieved. Both of these results significantly improve on existing manually designed features.