[8p-A23-1~12] CS.2 Code-sharing session of 3.2&4.4

3:00 PM - 3:15 PM JST(6:00 AM - 6:15 AM UTC)Presentation by Applicant for JSAP Young Scientists Presentation Award

[8p-A23-5] Fundamental Study on AI-Based Correction Process for AI-Assist 3D-Printed Lens

〇Akina Tsubuku¹, Hayato Morishita¹, Masanori Takabayashi^1,2, Atsushi Shibukawa³ (1.Kyutech, 2.Neumorph Center, Kyutech, 3.IPHF, Tokushima Univ.)

Keywords:

Deep Learning

This study investigates AI-based image correction to compensate for imperfections of 3D-printed lenses. While 3D printing enables low-cost, rapid fabrication of arbitrary lens shapes, surface roughness and shape errors degrade imaging quality. Two deep learning models, U-Net and CycleGAN, were applied to correct images captured through a 3D-printed convex lens in coherent imaging, using 1,577 paired images for training. U-Net achieved an average PSNR gain of 6 dB and SSIM improvement of 0.3. CycleGAN showed visible correction effects but was inferior to U-Net, with gains of 2 dB in PSNR and 0.2 in SSIM. U-Net provides superior performance, while CycleGAN is viable when paired data is unavailable. Future work will optimize both lens fabrication and AI correction processes.