In this project, we explore automated aesthetic evaluation of photographs using machine learning and image processing techniques.
We theorize that the spatial distribution of certain visual elements within a given image correlates with its aesthetic quality. To this end, we present a novel approach wherein we model each photograph as a set of tiles, extract visual features from each tile, and train a classifier on the resulting features along with the images’ aesthetics ratings.
Our model achieves a 10-fold cross-validation classification success rate of 83:60%, corroborating the efficacy of our methodology and therefore showing promise for future development.
Table 1 shows our 10-fold cross-validation accuracy for each of the learning algorithms. For both datasets, we got the highest performance with GBRTs, with accuracies of 80:88% and 83:60%. That we see similar quality of results for both datasets signifies that our methodology is sound.
Figure 3 shows the confusion matrix for 10-fold cross validation using GBRTs on the photo.net dataset. The true positive and false negative rates are approximately symmetric with the true negative and false positive rates, respectively, which signifies that our classifier is not biased towards predicting a certain class. This also holds true for the DPChallenge dataset.