Zero-Shot Haze Removal Using Near-Infrared and Color Image Fusion
Haze degrades outdoor images by reducing visibility, color fidelity, and edge contrast. This research develops a zero-shot haze removal method that combines near-infrared (NIR) and color images without requiring a large training dataset. The proposed method first restores color information from the visible image while extracting clear structural information from the NIR image using an atmospheric scattering model. The restored images are then combined through an edge-fusion process designed to compensate for luminance differences between the two modalities. This enables detailed NIR edge information to be incorporated while preserving natural color and brightness. Experimental results demonstrate improved color reproduction and sharper edge restoration compared with conventional methods, particularly under strong haze conditions.
Publication
Onhi Kato and Akira Kubota, “Zero-Shot Learning-Based Haze Removal Through Near-Infrared and Color Images Fusion,” IEICE Transactions on Information and Systems, vol. E109-D, no. 1, pp. 82–94, Jan. 2026. [Paper]

Perfect Reconstruction of Light Fields from Focal Stacks
Light fields represent both spatial and angular information of a scene, while a focal stack consists of images focused at different depths and can serve as an alternative representation. Reconstructing a light field from its focal stack is generally an ill-posed inverse problem. This research investigates conditions under which perfect reconstruction becomes possible. We show that, for Lambertian scenes without occlusion, a light field can be perfectly reconstructed from its focal stack using a filter bank when the camera aperture is modeled by a Cauchy function. Based on this property, perfect reconstruction filters are derived theoretically. Numerical simulations confirm that the proposed filter bank successfully reconstructs the original light field under the assumed conditions.
Publication
Akira Kubota, Kazuya Kodama, Daiki Tamura, and Asami Ito, “Filter Bank for Perfect Reconstruction of Light Field from Its Focal Stack,” IEICE Transactions on Information and Systems, vol. E106-D, no. 10, pp. 1650–1660, Oct. 2023. [Paper]

Spectral Reflectance Estimation of Facial Skin from a Single RGB Image
Spectral reflectance provides richer information about facial skin than conventional RGB values and enables accurate color reproduction under different illumination conditions. However, estimating a high-dimensional reflectance spectrum from only three RGB channels is inherently ill-posed. This research addresses the problem using a physics-based reconstruction framework with the expectation-maximization (EM) algorithm. The method models the image formation process using the spectral characteristics of the illumination and camera and iteratively estimates the skin reflectance from a single RGB image. By incorporating the physical relationship between spectral reflectance and RGB measurements, the proposed approach can recover plausible skin spectra without relying solely on a direct data-driven mapping. Experimental results demonstrate the effectiveness of the method for facial skin reflectance estimation.
Publication
Yoshihito Tanaka, Akira Kubota, and Shugo Yamaguchi, “Spectral Reflectance Estimation of Facial Skin from a Single RGB Image via EM-Based Physical Reconstruction,” 2026 IEEE International Conference on Image Processing (ICIP), 2026. (Top 3% Paper; Best Student Paper Award Finalist) 

Image Recoloring for Deutan Color Vision Deficiency with Fewer-Shot Self-Supervised Learning
Color vision deficiency can make it difficult to distinguish certain colors in everyday images. This research develops a self-supervised image recoloring method for Deutan color vision deficiency using only a small number of training images. The proposed method incorporates a Laplacian pyramid to separate an image into low- and high-frequency components. Recoloring is mainly applied to the low-frequency component, while high-frequency information is preserved to maintain edges and fine structures. The method also adapts the recoloring strength according to the severity of color vision deficiency, aiming to improve color discrimination while avoiding unnatural color changes. This framework enables effective recoloring with limited training data and preserves the visual naturalness of the original image.
Publication
Onhi Kato and Akira Kubota, “Fewer-Shot Self-Supervised Image Recoloring for Deutan Deficiency Based on Laplacian Pyramid,” 2025 IEEE International Symposium on Multimedia (ISM), pp. 21–22, 2025.

LPConv: Parameter-Efficient Receptive Field Expansion Using Laplacian Pyramids
Large receptive fields enable convolutional neural networks to capture global image structures, but simply increasing kernel size results in a substantial increase in the number of parameters. This research proposes LPConv, a parameter-efficient convolutional layer based on the Laplacian pyramid. LPConv decomposes feature maps into multiple resolutions and applies learnable convolutions mainly to low-frequency components, enabling effective receptive field expansion with fewer parameters. Experiments on ImageNet-1K show that LPConvNeXt-T achieves accuracy comparable to WTConvNeXt-T while reducing the parameter count by approximately one million. LPConv also demonstrates strong performance on CIFAR-100 and semantic segmentation on Pascal VOC 2012. Effective receptive field analysis further shows that LPConv produces smoother and wider receptive fields, providing a favorable balance between parameter efficiency and performance.
Publication
Naoki Nishiya and Akira Kubota, “LPConv: Laplacian Pyramid Convolutions for Parameter-Efficient Receptive Field Expansion,” 2026 IEEE International Conference on Image Processing (ICIP), 2026.