Signal Processing - July 2017 - 86

examples is around 1.07: 1. The Aesthetic and Attributes
considerable amount of engineering skill and domain experDatabase (AADB) [25] also contains a balanced distribution
tise. Next we review a variety of approaches that exploit
of professional and consumer photos, with a total of 10,000
hand-engineered features.
images. Eleven aesthetic attributes and annotators' IDs are
provided. A standard partition with 8,500 images for trainSimple image features
ing, 500 images for validation, and 1,000 images for testing
Global features are first explored by researchers to model the
is proposed [25].
aesthetic aspect of images. The works by Datta et al. [21] and
The trend toward creating data sets of even larger volume
Ke et al. [37] are among the first to cast aesthetic understandand higher diversity is essential for boosting the research
ing of images into a binary classification problem. Datta et
progress in this field of study. To date, the
al. [21] combine low-level and high-level
AVA data set serves as a canonical benchfeatures that are typically used for image
The distribution of positive retrieval and train an SVM classifier for
mark for performance evaluation of image
and negative examples in
aesthetic assessment, as it is the first largebinary classification of images in terms of
scale data set with detailed annotation.
aesthetic quality. Ke et al. [37] propose
the data set also plays a
Still, the distribution of positive and negarole in the effectiveness of global edge distribution, color distribution,
tive examples in the data set also plays a
hue count, and low-level contrast and
trained models.
role in the effectiveness of trained models,
brightness indicators to represent an
as false-positive predictions are as harmful
image; then they train a naïve Bayes clasas having a low recall rate in image retrieval and searching
sifier based on such features. An even earlier attempt by
applications. In the following, we review major attempts in the
Tong et al. [20] adopts boosting to combine global lowliterature to build systems for the challenging task of image
level simple features (blurriness, contrast, colorfulness,
aesthetic assessment.
and saliency) to classify professional photographs and ordinary snapshots.
Conventional approaches with handcrafted features
All of these pioneering works present the very first attempts
The conventional option for image quality assessment is to
to computationally model the global aesthetic aspect of imaghand-design good feature extractors, which requires a
es using handcrafted features. Even in a recent work, Ayd in

~160,000

~ 70,000

AVA Training Partition
~16,000

~4,000

AVA Testing Partition
Number of
Positive Images
Number of
Negative Images
(a)

(b)

FIGURE 4. Some sample images in the AVA data set [49]. (a) Images in the green-framed box are labeled with a mean score of >5. Images in the red-framed
box are labeled with a mean score of <5. The image groups on the right are ambiguous, with a somewhat neutral scoring around five. (b) The number of
images in the AVA data set.

IEEE SIGNAL PROCESSING MAGAZINE

July 2017

Table of Contents for the Digital Edition of Signal Processing - July 2017