Super-resolution light-field imaging based on multi-scale feature fusion

As a new generation of imaging equipment, a light-field camera can simultaneously capture the spatial position and incident angle of light rays. However, the recorded light-field has a trade-off between spatial resolution and angular resolution. Especially the limited spatial resolution of sub-aperture images limits the application scenarios of light-field cameras. Therefore, a light-field super-resolution network that fuses multi-scale features to obtain super-resolved light-field is proposed in this paper. The deep-learning-based network framework contains three major modules: multi-scale feature extraction module, global feature fusion module, and up-sampling module. The design ideas of different modules are as follows.

1) Multi-scale feature extraction module: To explore the complex texture information in the 4D light-field space, the feature extraction module uses ResASPP blocks to expand the perception field and to extract multi-scale features. The low-resolution light-field sub-aperture images are first sent to a Conv block and a Res block for low level feature ex-traction, and then a ResASPP block and a Res block are alternated twice to learn multi-scale features that accumulate high-frequency information in the 4D light-field.

2) Global feature fusion module: The light-field images contain not only spatial information but also angular information, which implies inherent structures of 4D light-field. The global feature fusion module is proposed to geometrically reconstruct the super-resolved light-field by exploiting the angular clues. It should be noted that the feature maps of all the sub-images from the upstream are first stacked in the channel dimension of the network and then are sent to this module for high-level features extraction.

3) Up-sampling module: After learning the global features in the 4D light-field structure, the high-level feature maps could be sent to the up-sampling module for light-field super resolution. This module uses sub-pixel convolution or pixel shuffle operation to obtain 2× spatial super-resolution, after feature maps are sent to a conventional convolution layer to perform feature fusion and finally output a super-resolved light-field sub-images array.

The network proposed in this paper was applied to the synthetic light-field dataset and the real-world light-field dataset for light-field images super-resolution. The experimental results on the synthetic light-field dataset and real-world light-field dataset showed that this method outperforms other state-of-the-art methods in both visual and numerical evaluations. In addition, the super-resolved light-field images were applied to depth estimation, and the results illustrated the parallax calculation enhancement of light-field spatial super-resolution, especially in occlusion and edge regions.

Fig. 1 Schematic of Optical field super-resolution net-work

About The Group

Associate Professor Shengxian Shi’s group mainly focuses on light field 3D metrology, light field volumetric flow diagnostic techniques, light field 3D reconstruction techniques. The researched are supported by NSFC, Gas Turbine State Key Grant, RC-NSFC. They have published more than 40 papers and won a series of awards like Shanghai Raising Star Programme Award, Tan Chin Tuan Exchange Fellowship.


Zhao Yuanyuan, Shi Shengxian. Light-field image super-resolution based on multi-scale feature fusion[J]. Opto-Electronic Engineering, 2020, 47(12): 200007.