Panoramic Light Field - Oriented Image Stitching Method with Semantics and Projection

Existing light field (LF) stitching approaches mainly consider the geometric correspondence and ignore the high-level semantic information, so it is difficult to ensure the angular consistency of the stitched results while aligning the LF images. To address this problem, this paper proposes a panoramic LF - oriented image stitching method with semantics and projection transformation. Specifically, semantic information is first extracted from the LF images and applied to classify feature points. Hence, all the feature points are endowed corresponding semantic information based on the classification result. Afterwards, the geometric and semantic information of the feature points are combined to estimate the local homography matrix to align the input LF images. Finally, a weighted fusion strategy is employed to fuse the aligned results. Experimental results show that the proposed method can not only guarantee the spatial quality of the stitched LF image. But also effectively preserve accurate angular consistency. Moreover, the proposed method outperforms the existing stitching methods in both subjective and objective comparisons.