Depth estimation method based on light-field data distribution

1. A depth estimation method based on light-field data distribution, comprising the following steps:

S 1 . adjusting pixel distribution of input light-field images to generate a series of refocused light-field images having different focal lengths;

S 2 . for a spatial point, corresponding to a macro-pixel, extracting intensity ranges for the macro-pixel from the refocused light-field images, and then selecting one refocused light-field image corresponding to a minimum intensity range, the focal length of the selected refocused light-field image being taken as a scene depth of the macro-pixel, the macro-pixel corresponding to one point in an actual scene, and the intensity range of the macro-pixel being a variation range of intensity values of all points within the macro-pixel; and

repeating the step S 2 to obtain scene depths of all macro-pixels.

2. The depth estimation method based on light-field data distribution according to claim 1 , wherein in the step S 1 , a point spread function is adopted to adjust the pixel distribution of input light-field images.

3. The depth estimation method based on light-field data distribution according to claim 1 , further comprising a step S 3 of performing, according to a confidence model, global optimization on the scene depths obtained in the step S 2 .

4. The depth estimation method based on light-field data distribution according to claim 3 , wherein the step S 3 of performing, according to the confidence model, the global optimization on the scene depth obtained in the step S 2 comprises: by taking the scene depth obtained in the step S 2 as an initial input, using a Markov random field to perform optimization, and a specific optimization method comprises: evaluating the depth of each point according to the confidence model, using a depth estimation having high accuracy to amend an inaccurate depth, improving the consistence of the depth estimation in homogeneous regions, and reserving a depth boundary.

5. The depth estimation method based on light-field data distribution according to claim 3 , wherein the confidence model is a multivariate confidence model, and the multivariate confidence model comprises a first portion for measuring accuracy of the scene depths, and a second portion for measuring consistency of the scene depths at non-boundary regions and mutability at boundary regions.

6. The depth estimation method based on light-field data distribution according to claim 5 , wherein

the first portion of the multivariate confidence model is C1(x, y),

C 1 ( x,y )=√{square root over (( R z* ( x,y )− R z′ ( x,y )) 2 +( z*−z ′) 2 )}

wherein R z* (x, y) and R z (x, y) are a minimum value point and a minimal value point of a changing curve of an intensity range R z (x, y) with a scene depth respectively, and z* and z′ are scene depths corresponding to the minimum value point and the minimal value point respectively.

7. The depth estimation method based on light-field data distribution according to claim 5 , wherein

the second portion of the multivariate confidence model is based on gradient information of central sub-aperture texture images;

the depth estimation method further comprises a step of acquiring central sub-aperture texture images of the series of refocused light-field images respectively, and a step of using the acquired central sub-aperture texture images to calculate the credibility by the second portion of the multivariate confidence model.