Fig 1.
Image blurring in imaging system.
Fig 2.
Examples of the PSF (first row) and its frequency domains (second row).
Left: Motion blur with (L,Φ) = (30,45°) in Eq (2). Middle: Defocus blur with r = 9 in Eq (3). Right: Gaussian blur with σ = 3.5 in Eq (4).
Fig 3.
Examples of the blur feature in motion blur.
Left: Motion blur of the “bikes.bmp” image with L =10 (first row) and L = 20 (second row) and orientation Φ = 120°. Middle: Gradient domain. Right: Amplitudes of the blur features.
Fig 4.
Examples of the blur feature in defocus blur.
Left: Defocus blur of the “monarch.bmp” image with r = 3 (first row) and r = 3 (second row). Middle: Gradient domain. Right: Amplitudes of the blur features.
Fig 5.
Examples of the blur feature in atmospheric turbulence blur.
Left: Atmospheric turbulence blur of the “house.bmp” image with σ = 2.0 (first row) and σ = 4.8 (second row). Middle: Gradient domain. Right: Amplitudes of the blur features.
Fig 6.
Basic structure of a GRNN.
Fig 7.
Framework of the proposed method.
The “Estimation” and “Deblurring” processes are described in Algorithms 1 and 2, respectively.
Fig 8.
Examples of the pascal VOC dataset.
Fig 9.
Cost duration for different p values. Algorithm 2 was tested for a 768×512×3 image that was blurred by a series of defocus blurs. The x-axis represents the radius while the y -axis shows the runtime (in seconds).
Fig 10.
Examples of the ground truth images.
(A) “caps.bmp.” (B) “lighthouse2.bmp.” (C) “monarch.bmp.” (D) “plane.bmp.”.
Table 1.
Test images used and their blur parameters.
Fig 11.
Motion blur with L = 18 and Φ = 30° for the “caps.bmp” image.
(A) Blurred image. (B) Method in [49]. (C) Method in [50]. (D) Method in [51]. (E) Method in [52]. (F) Proposed method.
Table 2.
IQA data for the comparison results of Figs 11, 12, 14 and 16.
Fig 12.
Motion blur with L = 16 and Φ = 45° for the “lighthouse2.bmp” image.
(A) Blurred image. (B) Method in [49]. (C) Method in [50]. (D) Method in [51]. (E) Method in [52]. (F) Proposed method.
Fig 13.
Results of comparisons of test motion-blurred images.
The corresponding motion length, L, and orientation, Φ, values are listed in the third column of Table 1.
Fig 14.
Defocus blur with r = 4 for the “monarch.bmp” image.
(A) Blurred image. (B) Method in [49]. (C) Method in [50]. (D) Method in [51]. (E) Method in [52]. (F) Proposed method.
Fig 15.
Results of comparisons of test defocus-blurred images.
The corresponding radius, r, values are listed in the fourth column of Table 1.
Fig 16.
Atmospheric turbulence blur with σ = 3.2 for the “plane.bmp” image.
(A) Blurred image. (B) Method in [49]. (C) Method in [50]. (D) Method in [51]. (E) Method in [52]. (F) Proposed method.
Fig 17.
Results of comparisons of test atmospheric turbulence-blurred images.
The corresponding parameter, σ, values are listed in the fifth column of Table 1.
Fig 18.
Results of average processing speed for the compared methods.
The x-axis lists the PSF sizes, which range from 3×3 to 25×25, while the y-axis shows the total runtime (in seconds).
Fig 19.
Motion-blurred photograph in real-life application.
Fig 20.
Results of comparisons with Fig 19.
(A) Blurred image. (B) Method in [49]. (C) Method in [50]. (D) Method in [51]. (E) Method in [52]. (F) Proposed method.