Learning to Sharpen Partially Blurred Image via Iterative Blurred Region Mining and Recovery

Despite progress on image deblurring, existing advanced deep-net models are difficult to directly process partially blurred images due to the mismatch in the distribution of training (full-scaled blurred) and test (partially blurred) images. To address this problem, we presents a novel iterative blurred region mining and recovery (iBRMR) approach, which can automatically and progressively mine and sharpen the blurred regions. Our iBRMR approach can handle not only partial blur but also full-scaled blur, and shows superior deblurring performance.

Video demo

The deblurring results are all done by our iBRMR approach.

Partially Blurred GOPRO Dataset:

Single Blurred Region

1blurRegion-peak.mp4

Original partially blurred image iBRMR-modified VGG16 (Ours)

The video is best viewed in 1080P (HD). Please set it from playback options.

Partially Blurred GOPRO Dataset:

Two Blurred Regions

2blurRegion-peak.mp4

Original partially blurred image iBRMR-modified VGG16 (Ours)

The video is best viewed in 1080P (HD). Please set it from playback options.

GOPRO Dataset [1]:

Full-scaled Blurred Images


full-scale.mp4

Original full-scaled blurred image iBRMR-modified VGG16 (Ours)

The video is best viewed in 1080P (HD). Please set it from playback options.

Some other results

(from Posetrack dataset [2])

Blurred image iBRMR-modified VGG16 (Ours)

Partially Blurred GOPRO Dataset:

  • Single blurred region: A total of 1,111 pairs of test data are generated, each pair contains a partially blurred image and a corresponding sharp image.

  • Two blurred regions: Similarly, we generated a total of 1,111 pairs of test data, where each pair contains an image with two blurred regions and a corresponding sharp image.

The dataset can be downloaded via the link below:

https://drive.google.com/file/d/1FyJkJSN5-ZLP0ozhhgf1EN-HBQLPaM6t/view?usp=sharing

Here are some example images.

Single blurred region image

Two blurred regions image

Sharp image

Reference

[1] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2017.

[2] Mykhaylo Andriluka, Umar Iqbal, Eldar Insafutdinov, Leonid Pishchulin, Anton Milan, Juergen Gall, and Bernt Schiele. Posetrack: A benchmark for human pose estimation and tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018.