(Weights) Image Patches CNN × CNN × CNN × Shared Model Statistical Moments CNN + Regression CNN CNN Pooled Score Predicted Scores Image Patches Shared Model (a) Feature Vectors (b) (Weight Maps) CNN NN CNN × CNN Image Feature Maps Aggregated Feature Vector Σ Regression Aggregated Feature Vector Regression CNN Image Patches Shared Model (c) Feature Vectors Aggregated Feature Vector (d) FIGURE 4. Examples of aggregation and pooling strategies in CNN-based picture-quality prediction models. (a) Weighted average pooling, (b) elementwise aggregation, (c) weighted average aggregation, and (d) NN aggregation. JPEG2000 GB WN Distortion Severity Selected Distortion Parameters Distortion Simulation Pristine Image (a) * Original Scene * Shooting Environment * Shooting Device * User Skill Sensor to Storage Pipeline Demosaicing Denoising Compression... Mixtures of Image Distortions (b) FIGURE 5. (a) Synthetic and (b) authentic image distortions found in picture-quality databases. IEEE SIGNAL PROCESSING MAGAZINE | November 2017 | 137