Assessing Image Aesthetics Using Machine Learning

Abstract

Automatically assessing an image's visual aesthetic quality is a challenging problem with many potential applications, including better photo management and improvement of skills such as photography and handwriting. We focused on classifying the aesthetics of two types of images. The first category consisted of naturally taken photographs of people, buildings, nature, objects, etc. The second category consisted of images of handwritten digits. The classifiers were trained using aesthetic quality ground truth labels based on people's majority votes. For classifying natural photographs as having either good or bad aesthetic quality, we used a radial basis function (RBF) support vector machine (SVM) with features based on those proposed by prior research to achieve an accuracy of 84.8%. For evaluating the aesthetics of handwritten digits, we performed a variety of experiments with SVMs and convolutional neural networks (CNNs). For binary \good" versus \bad" handwriting aesthetic quality classification, our RBF SVM classifier achieves an accuracy of 75.1% and our CNN classifier achieves an accuracy of 81.0%. This is lower than but comparable to the average accuracy of a human evaluator, which is 88.2%. When limiting training and testing to a particular digit, accuracy generally increased for both the SVM and CNN classifiers, reaching up to 89%. We demonstrated that when the number of training examples is sufficiently large, CNNs classify the aesthetics of handwritten digits with higher accuracy than SVMs. These handwriting aesthetics classification experiments lay the groundwork for future research on creating an aid to help diagnose disorders associated with handwriting difficulties.