Co-optimized tunable-focus light field imaging system for 3D seed phenotyping: From optical design to computational reconstruction

Three-dimensional seed phenotyping demands imaging systems that simultaneously achieve the micron-level resolution in centimeter-level field-of-view (FOV), a challenge exacerbated by the intrinsic resolution-FOV trade-off in conventional light field architectures. This paper presents a co-optimized framework integrating a dynamically reconfigurable optical system with computational imaging pipelines, to meet the demand from the variety of seed phenotyping research. At the hardware level, we develop a tunable-focus lens group containing main lens with tunable lens, enabling the flexible adjustment on the effective focal length, coupled with a custom microlens array. A mathematic model is analyzed with FOV, lateral resolution, DOF, lens parameters and system configurations, which also help us on individual optical component design (i.e., MLA design). Computationally, we propose a hybrid aberration correction strategy: First, co-calibration of lens and microlens array aberrations via line-feature detection is developed. Subsequently, a conditional generative adversarial network (cGAN) with attention-guided residual learning enhances sub-aperture images, attaining PSNR 34.63 dB and SSIM 0.9570 on seed test. Experimentally, system achieves 6.2lp/mm resolution at MTF50 over 2~3cm FOV, a 307% improvement over 1.52 lp/mm at initial configurations. The reconstruction pipeline synergizes epipolar plane image (EPI) analysis with multi-view consistency constraints from the sub-aperture array, generating dense 3D point clouds surface (approximately 1.5×10^4 points/cm²) that preserve spectral-textural features at the mean time. Experimental validation of bitter melon seeds and rice grain seeds demonstrates that the accurate morphological parameters are extracted in large area and present in high-fidelity 3D reconstruction. This hardware/software co-optimization framework demonstrates unprecedented dynamic adjustability of resolution-FOV trade-off, overcoming the inherent limitations of conventional light-field systems, while establishing a field-reconfigurable scalable architecture for next-generation phenotyping, with potential extensions to robotic vision and biomedical imaging applications.