研究目的
To address the challenging problem of defocus blur detection from a single image by applying a local sharpness metric, which obtains from the CNN-based of feature learning in the blur and non-blur image regions.
研究成果
The proposed algorithm achieves state-of-the-art performance on defocus blurred image detection by combining the advantages of both ConvNets and local metric, and introducing a novel iterative updating mechanism to refine the detection results.
研究不足
The method may have challenges in precisely differentiating an in-focus smooth region and a blurred smooth region. The iterative sharpness metric detection part consumes significant computation time.
1：Experimental Design and Method Selection:
The methodology involves training convolutional neural networks (ConvNets) to learn image features at the super-pixel level, extracting convolution kernels, and applying principal component analysis (PCA) to obtain a local sharpness metric. An iterative updating mechanism using the hyperbolic tangent function refines the detection results.
2：Sample Selection and Data Sources:
A public blurred image dataset consisting of 704 partially blurred images and accompanying hand-segmented ground truth images was used. Patches were extracted around super-pixels for training.
3：List of Experimental Equipment and Materials:
Intel Core i5-4200 desktop computer, equipped with 2.50 GHz dominant frequency and 8 GB RAM. Software includes TensorFlow for training ConvNets and Matlab for run-time tests.
4：50 GHz dominant frequency and 8 GB RAM. Software includes TensorFlow for training ConvNets and Matlab for run-time tests.
Experimental Procedures and Operational Workflow:
4. Experimental Procedures and Operational Workflow: The process includes training ConvNets on blur and sharp patches, extracting and processing convolution kernels with PCA, and refining detection results iteratively.
5：Data Analysis Methods:
Precision and recall curves were generated by varying the threshold used to produce a segmentation of the final sharpness maps.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
