Signal Processing - July 2017 - 92

Table 2. The methods evaluated on the CUHK-PQ data set.
Method

Data Set

Metric

Result

Training-Testing Remarks

Su et al. (2011) [72]

CUHK-PQ

Overall accuracy

92.06%

1,000 training, 3,000 testing

Marchesotti et al. (2011) [47]

CUHK-PQ

Overall accuracy

89.90%

50-50 split

Zhang et al. (2014) [67]

CUHK-PQ

Overall accuracy

90.31%

50-50 split, 12,000 subset

Dong et al. (2015) [50]

CUHK-PQ

Overall accuracy

91.93%

50-50 split

Tian et al. (2015) [54]

CUHK-PQ

Overall accuracy

91.94%

50-50 split

Zhang et al. (2016) [57]

CUHK-PQ

Overall accuracy

88.79%

50-50 split, 12,000 subset

Wang et al. (2016) [53]

CUHK-PQ

Overall accuracy

92.59%

4:1:1 partition

Lo et al. (2012) [66]

CUHK-PQ

Area under ROC curve

0.93

50-50 split

Tang et al. (2013) [45]

CUHK-PQ

Area under ROC curve

0.9209

50-50 split

Lv et al. (2016) [51]

CUHK-PQ

Mean AP

0.879

50-50 split

Table 3. The methods evaluated on the AVA data set.
Method

Data Set

Metric

Result

Training-Testing Remarks

Marchesotti et al. (2013) [48]

AVA

ROC curve

tpr: 0.7, fpr: 0.4

Standard partition

AVA handcrafted features (2012) [49]

AVA

Overall accuracy

68.00%

Standard partition

Spatial pyramid pooling (SPP) (2015) [24]

AVA

Overall accuracy

72.85%

Standard partition

RAPID (full method) (2014) [23]

AVA

Overall accuracy

74.46%

Standard partition

Peng et al. (2016) [52]

AVA

Overall accuracy

74.50%

Standard partition

Kao et al. (2016) [58]

AVA

Overall accuracy

74.51%

Standard partition

RAPID (improved version) (2015) [55]

AVA

Overall accuracy

75.42%

Standard partition

DMA-net (2015) [24]

AVA

Overall accuracy

75.41%

Standard partition

Kao et al. (2016) [59]

AVA

Overall accuracy

76.15%

Standard partition

Wang et al. (2016) [53]

AVA

Overall accuracy

76.94%

Standard partition

Kong et al. (2016) [25]

AVA

Overall accuracy

77.33%

Standard partition

BDN (2016) [56]

AVA

Overall accuracy

78.08%

Standard partition

Zhang et al. (2014) [67]

AVA

Overall accuracy

83.24%

10% subset, 12.5k*2

Dong et al. (2015) [50]

AVA

Overall accuracy

83.52%

10% subset, 19k*2

Tian et al. (2016) [54]

AVA

Overall accuracy

80.38%

10% subset, 20k*2

Wang et al. (2016) [53]

AVA

Overall accuracy

84.88%

10% subset, 25k*2

Lv et al. (2016) [51]

AVA

Mean AP

0.611

10% subset, 20k*2

indication of 0.5 # (14k/14k) + 0.5 # (0k/6k) = 50% performance on AVA.
In this regard, in the following sections where we discuss
our findings on a proposed strong baseline, we report both
overall classification accuracy and balanced accuracy to get a
more reasonable measure of baseline performance.

Experiments on deep-learning settings
It is evident from Table 3 that deep learning-based approaches
dominate the performance of image aesthetic assessment.
The  effectiveness of learned deep features in this task has
92

motivated us to take a step back to consider how a CNN
works to understand the aesthetic quality of an image. It is
worth noting that training a robust deep aesthetic scoring
model is nontrivial, and often we found that the devil is in the
details. To this end, we design a set of systematic experiments
based on a baseline one-column CNN and a two-column
CNN, and evaluate different settings from minibatch formation to complex multicolumn architecture. The results are
reported on the widely used AVA data set.
We observe that by carefully training the CNN architecture, the two-column CNN baseline reaches comparable or

IEEE SIGNAL PROCESSING MAGAZINE

|

July 2017

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