richzhang/PerceptualSimilarity
PerceptualSimilarity
PerceptualSimilarity is a deep learning framework designed to quantify and evaluate the perceptual distance between images. It provides a system for measuring how similar two images or image patches appear to human vision by using deep feature representations instead of pixel-wise differences.
The project implements a differentiable distance metric that functions as a loss function, allowing image pixels to be optimized via backpropagation to reach a target visual appearance. It includes a trainable linear layer that can be applied to frozen deep features to learn weighted distance metrics aligned with human perception.
The framework covers broad capabilities in image quality evaluation, similarity metric training, and computer vision benchmarking. Model accuracy is evaluated by comparing predicted distance scores against human judgment datasets using frameworks such as two-alternative forced choice tests.
Features
- Perceptual Loss - Provides a differentiable perceptual loss function based on feature similarity for visual quality optimization.
- Computer Vision Models - Provides a deep learning framework for training and evaluating models focused on perceptual similarity.
- Deep Feature Distances - Implements similarity calculations based on Euclidean distances between high-level deep feature maps.
- Metric Learning - Trains linear layers on top of image representations to create custom metrics for visual similarity.
- Perceptual Metric Heads - Uses a trainable linear layer on top of frozen features to learn weighted human-perceptual distances.
- Perceptual Metrics - Calculates the distance between images based on human visual perception using deep feature representations.
- Image Quality Assessment - Compares processed images against originals to evaluate if visual artifacts are noticeable to humans.
- Perceptual Similarity Scoring - Calculates numerical values to quantify the perceived difference between images using deep learning features.
- Visual Similarity Training - Allows creating linear layers on image representations or optimizing models using specific visual training datasets.
- Computer Vision Benchmarks - Tests the accuracy of visual similarity models against standardized human judgment datasets.
- Iterative Image Optimizers - Implements iterative gradient descent to refine image pixels by minimizing perceptual distance.
- Two-Alternative Forced Choice Tests - Evaluates model accuracy by comparing predicted distance scores against human preferences in a binary choice framework.
- Perceptual Image Optimization - Adjusting the pixels of an image by using perceptual distance as a loss function to reach a target visual look.
- Perceptual Model Evaluation - Tests similarity models against standard datasets using human judgment modes like just-noticeable difference.
- Computer Vision Research - Deep features as a metric for perceptual image similarity.
- Evaluation Metrics - Learned perceptual image patch similarity metric.
- Image Transformation - Deep features as a perceptual metric.
