Research On Light Field Compression

Light field can fully describe all the information of the light, which will bring revolutionary changes to the traditional image processing methods. The element of traditional cameras on the imaging plane is purely the sum of the rays from different orientations. However, light field image’s information contains the direction of rays. It makes it easier to trace the ray. Since light field image has higher dimension and contains more information, the file size of the light field image is much larger than tradition images, which brings forth challenges to the storage and transmission band width.Therefore, the research of the light field image compression becomes necessary.This thesis made researches on several commonly used methods of light field image compression. Two main research fields were determined, i.e. light field image compression based on the disparity compensated and video coding methods respectively. This thesis put forward new technologies to improve the compression ratio of light field image compression based on the above two methods. The main contributions are as follows:First, we proposed a light field image compression method based on perspective transformation. When using one light field sub-image to predict another, we adopted perspective transformation to improve the correlation between these two images. Since the perspective transformation may bring some prediction error, we chose reference image with the rate-distortion criterion to improve prediction result.Second, we proposed a light field image compression method based on disparity compensation which uses 2D disparity. We adopted 2D disparity to eliminate the errors caused by the light field capture devices’ coupling and calibrating. Experiment results showed, with small sacrificing in file size, the reconstructed images have better qualities compared to those which compressed by traditional methods.Third, we proposed a method of optimizing the reference structure in the group of picture(GOP). We selected a better key sub-image in each GOP, ensuring that it has high correlation with other sub-images. We optimized the reference structure according to the correlation between each light field sub-images. We used bi-directional prediction to compress the light field sub-image. For each current image, there are two reference pictures. We chose the reference image which has the lowest Rate-Distortion(R-D)cost.