Iterative Filter Adaptive Network (IFAN)

Our deblurring network is built upon a simple encoder-decoder architecture consisting of a feature extractor, reconstructor, and IFAN module in the middle. The feature extractor extracts defocused features and feeds them to IFAN. IFAN removes blur in the feature domain by predicting spatially-varying deblurring filters and applying them to the defocused features using IAC. The deblurred features from IFAN is then passed to the reconstructor, which restores an all-in-focus image.

Iterative Adaptive Convolution Layer

The IAC layer iteratively computes feature maps as follows (refer Eq. 1 in the main paper):

Separable filters in our IAC layer play a key role in resolving the limitation of the FAC layer. Our IAC layer secures larger receptive fields at much lower memory and computational costs than the FAC layer by utilizing 1-dim filters, instead of 2-dim convolutions. However, compared to dense 2-dim convolution filters in the FAC layer, our separable filters may not provide enough accuracy for deblurring filters. We handle this problem by iteratively applying separable filters to fully exploit the non-linear nature of a deep network. Our iterative scheme also enables small-sized separable filters to be used for establishing large receptive fields.

Disparity Map Estimation & Reblurring

To further improve the single image deblurring quality, we train our network with novel defocus-specific tasks: defocus disparity estimation and reblurring.

Disparity Map Estimation exploits dual-pixel data, which provides stereo images with a tiny baseline, whose disparities are proportional to defocus blur magnitudes. Leveraging dual-pixel stereo images, we train IFAN to predict the disparity map from a single image so that it can also learn to more accurately predict blur magnitudes.

Reblurring, motivated by the reblur-to-deblur scheme, utilizes deblurring filters predicted by IFAN for reblurring all-in-focus images. For accurate reblurring, IFAN needs to predict deblurring filters that contain accurate information about the shapes and sizes of defocus blur. Based on this, during training, we introduce an additional network that inverts predicted deblurring filters to reblurring filters, and reblurs an all-in-focus image.

The Real Depth of Field (RealDOF) test set

We present the Real Depth of Field (RealDOF) test set for quantitative and qualitative evaluations of single image defocus deblurring. Our RealDOF test set contains 50 image pairs, each of which consists of a defocused image and its corresponding all-in-focus image that have been concurrently captured for the same scene, with the dual-camera system. Refer Sec. 1 in the supplementary material for more details.