Fast Direct Super-Resolution by Simple Functions Chih-Yuan Yang and Ming-Hsuan Yang 12/14/13
Outline Introduction Related work Proposed method Experimental results Conclusions
Introduction Super-Resolution = Predicting pixel intensities From single or multiple image(s) For spatial or temporal intestines 480 120 320 80
Challenges Effectiveness Stability Computational load Numerous pixel intensities Stability Various image content Computational load
Related Work - Bicubic Interpolation Predicting intensities by interpolating nearby pixels Simple and fast But over-smooth
Self-Similar Exemplars [Glasner 11] Predict HR patches through example patches found in a self-generated image pyramid Sharp and clear edges Blurred textures Slow I1 I0 I-1 I-3 I-2 I2 I3 I4 ICCV
Gradient Profile Prior [Sun08] Predict intensities by a edge model Sharp edges, but jaggy Only effective for edges
Sparse Representation [Yang08] Predicting HR patch features by learned sparse dictionaries Rich details Noise artifacts along edges
Proposed Method Split LR feature space for effectiveness Use simple features and simple functions for speed Exploit a large image set to collect training samples Asymmetric computational load slow for training fast for test
Training Phase (1) Generate LR images Generate LR images from HR training images 𝐼 𝑙 = 𝐼 ℎ ⊗𝐺 ↓ 𝑠 Extract patch pairs in LR and HR Since there is a convolution for HR images, we crop LR patches affected by predicted HR pixels (blue region)
Training Phase (2) Extract LR/HR features LR: LR intensity minus the mean of the LR patches HR: HR intensity minus the mean of the LR patches High-frequency information Cluster the LR training features by K-mean split the LR feature space into K subspaces
Cluster Centers and Patch Numbers
Training Phase (3) Collect training instances Randomly select sufficeint LR/HR feature pairs for each cluster For some cluster centers, use bicubic interpolation if no sufficient instances available
Training Phase (4) Learn regression coefficients C: coefficients W: HR features (1000 instances) V: LR features (1000 instances)
Test Phase Crop LR patches Compute patch mean, and extract LR features Find the closest cluster center Compute HR features HR patch intensity = HR feature + LR patch mean Average overlapped HR patches as output
Experimental Results - Child
Experimental Results - Lena
Experimental Results - IC No groundtruth image
Experimental Results - Mansion
Experimental Results - Mermaid
Experimental Results - Shore
Experiments on BSD200 dataset
Performance (averaged) and Execution Time for BSD200 Dataset
Conclusions Clear and sharp edges Rich texture details Easy for implementation Fast to generate SR images
Insight Split of the LR feature space K has to be large enough Exploit a large image set to collect sufficient examples for high-frequency patches Regression methods make little difference
Code and Dataset Available https://eng.ucmerced.edu/people/cyang35 Including code for training phase