G oal: Assess feasibility of developing an aesthetic label classifier for abstract images generated by RMIT’s Imagene software. VS.

Classification of the aesthetic value of images based on histogram features

By Xavier Clements & Tristan PenmanSupervisors: Vic Ciesielski, Xiadong Li

Acknowledgment: Rahayu Binti A Hamid

Goal: Assess feasibility of developing an aesthetic label classifier for abstract images generated by RMIT’s Imagene software.

VS.

Interesting Not Interesting

Image Classification

• Classification Algorithms– Support Vector Machines

– Random Committee ensemble algorithm• Random Forest Base classifier

Computational Aesthetics / Histogram Features

• Computational Aesthetics– is the analysis of image sets for their aesthetic value.– Yeowen Wu et al study - “The good, the bad, and the

ugly: Predicting aesthetic image labels” - 2010• Histograms– The use of histogram features was chosen as a way

of attaining a global description of each image, this method having been employed successfully in previous studies (Chapelle et al 1999).

Software

• Image Generation – RMIT’s Imagene System

• Feature Extraction– GNU Image Finding Tool (GIFT)

• Data Mining– WEKA 3 Data Mining Software suite (WEKA)• Sequential Minimum Optimization (SMO) algorithm• Random Committee with Random Forest base

(RCRF)classifier

Methodology1. Generate 5 Image test sets with RMIT’s Imagene software.

Move each image into either interesting or not interesting directories.

2. Extract features from each image via the GNU Image Finding Tool (GIFT).

3. Unpack binary feature files for each image and merge them into a feature matrix (CSV).

4. Cut down feature matrix to the colour and Gabor histogram attributes.

5. Import histogram feature matrix into WEKA and train SMO and Random Committee classifiers via 10-fold Cross Validation.

Results – Imagene Images

Results / Conclusions• Random Committee with Random Forest Base Classifier

(RCRF) - 94.52%• SMO - 93.57% • RCRF outperformed the SMO in overall classification, as

well as having higher precision and recall values for the interesting class.

• Conclusion: – The higher than expected classification accuracies ensure that

a classifier (RCRF or SMO) can be used to delineate relatively accurately between interesting and not interesting Imagene images.