Fig. 1. Framework of single image dynamic refocusing algorithm

Fig. 2. Light field digital refocusing. (a) Sensor image; (b) sub-aperture image; (c) refocused image

Fig. 3. Principle of CoC rendering by gathering method

Fig. 4. Deep network architecture. (a) Focused stack based method; (b) parallax based method

Fig. 5. Comparison of monocular depth estimation results. (a) Center image; (b) disparity method; (c) focal stack method; (d) method proposed by Godard et al.^[16]; (e) normalized method proposed by Cheng et al.^[17](SSIM:0.756); (f) normalized disparity method (SSIM:0.823); (g) normalized focal stack method (SSIM: 0.895); (h) method proposed by Jeon et al.^[18]

Fig. 6. Quantitative and qualitative comparison of light field synthesis methods. (a) Method proposed by Kalantari et al.^[10]; (b) method proposed by Srinivasan et al.^[13]; (c) our method; (d) ground truth center image; (e)(f)(g) rendered center images obtained by methods proposed by Kalantari et al.^[10] and Srinivasan et al.^[

Fig. 7. Quantitative and qualitative comparison of rendering methods. (a) Center image; (b1)--(b4) refocused images; (c) depth estimation using focal stack method; (d1)--(d4) rendering results using focal stack method; (e) depth estimation using disparity method; (f1)--(f4) rendering results using disparity method; (g)(h)(i) results of occlusion detection, focus on foreground, and focus on background obtained by method proposed by Zhang et al.^[5]; (j)(k)

Fig. 8. Comparison of rendering results on different datasets. (a) Ground truth center image; (b) depth estimation; (c) refocusing on close positions; (d) refocusing on far positions

Fig. 9. Comparison of rendering effects of real scenes with images captured by different cameras. (a) Original image; (b)--(e) refocused images at four depths rendered by our method;(f) depth map; (g)(h) images shot by dual cameras focused on two positions; (i)(j) images shot by Cannon focused on two positions