Automatic Extraction of Blur Regions on a Single Image Based on Semantic Segmentation

Aodong Shen, Han Dong, Kun Wang, Youyong Kong, Jiasong Wu, Huazhong ShuIEEE Access 2020 · DOI

Introduction

Identifying spatially-varying blur—where only specific regions of an image are out of focus or distorted by motion—is a fundamental challenge in computer vision. Accurate blur maps are essential for downstream tasks such as object detection and image restoration. Historically, researchers relied on hand-crafted features like gradient distributions or local binary patterns. However, these approaches are often brittle, requiring significant domain expertise and failing to account for global context. For instance, a smooth, textureless wall might be incorrectly classified as blurry simply because it lacks sharp edges.

Methodology: The BDNet Framework

To overcome these limitations, researchers developed BDNet (Blur Detection Net), a deep learning framework that treats blur detection as a semantic segmentation task. The architecture utilizes a "U-shaped" encoder-decoder structure:

  • Encoder: A ResNet backbone extracts high-level semantic features from the input image. By leveraging a pre-trained model via transfer learning, BDNet can utilize general visual knowledge even with limited blur-specific data.
  • Decoder: A Fully Convolutional Network (FCN) processes these features to generate a pixel-level binary mask, distinguishing blur from non-blur regions.
  • Skip Connections: The model fuses low-level spatial details from early layers with high-level semantic information from deeper layers to preserve boundary precision.
  • Optimization: Unlike traditional FCNs trained in stages, BDNet is trained end-to-end. The inclusion of Batch Normalization before feature merging stabilizes training and enhances accuracy.

Datasets and Training

The model was evaluated using three primary datasets derived from 1,000 annotated images (296 motion blur, 704 defocus):

  1. BlurRaw: The original foundation dataset.
  2. BlurDB1: 7,072 random 256x256 patches.
  3. BlurDB2: 9,134 patches generated via a sliding window with 50% overlap to maximize training data.

The researchers found that joint training of the encoder and decoder with a low learning rate yielded superior results compared to freezing the encoder, suggesting that blur detection requires specialized feature refinement.

Performance Evaluation

BDNet demonstrates a significant performance leap over traditional algorithms and general-purpose segmentation networks like SegNet and UNet.

Examples of blur detection tasks showing original images (top) and their corresponding binary masks (bottom).

The model achieved a Mean IoU of 0.800, representing a nearly 20% increase over the best traditional hand-crafted methods (which range from 0.329 to 0.584). Both Pixel Accuracy and Mean Accuracy exceeded 90%. By maintaining a large receptive field through its fully convolutional structure, BDNet successfully uses surrounding context to identify textureless sharp regions that would otherwise confuse local patch-based models.

Conclusion and Future Work

While BDNet is robust across different blur types (motion vs. defocus), challenges remain. Very large, smooth areas, such as clear skies, can still be ambiguous due to a lack of gradients. Additionally, the "soft" boundaries inherent in motion blur are more difficult to delineate precisely compared to the sharper boundaries often found in defocus blur. Future research will explore incorporating more global context and utilizing higher-resolution training data to refine these edge cases.

Summary updated

BibTeX
@article{shen2020automatic,
  author = {Aodong Shen and Han Dong and Kun Wang and Youyong Kong and Jiasong Wu and Huazhong Shu},
  title = {{Automatic Extraction of Blur Regions on a Single Image Based on Semantic Segmentation}},
  journal = {IEEE Access},
  year = {2020},
  doi = {10.1109/access.2020.2978084},
}
Figure