Jong-Ok Kim

Near-infrared (NIR) imaging has been widely utilized in low-light applications due to its ability to capture fine textures with less noise compared to RGB sensors, but the absence of color information makes NIR imagery less intuitive for human perception. To bridge this perceptual gap, NIR colorization aims to translate NIR images into visually interpretable RGB representations. In particular, multi-band NIR colorization has attracted growing attention for its potential to leverage complementary spectral information across multiple wavelengths, enabling more accurate NIR-to-RGB mappings. Despite this progress, the inherent domain gap between multi-band NIR and RGB introduces structural ambiguities that hinder accurate reconstruction. To address this issue, we propose a two-stage NIR-to-Visible (N2V) colorization module designed to progressively align NIR representations with the visible domain. In the first stage, a Distance-Assisted Deformable Convolution (DA-DCN) module enhances spatial adaptability to effectively capture fine-grained patterns and structural details within the NIR domain. In the second stage, a Deformable Distillation mechanism transfers spatial sampling priors from a teacher network with auxiliary grayscale input to a student network that relies solely on multi-band NIR data, thereby facilitating domain alignment and suppressing visual artifacts. Extensive experiments demonstrate that the proposed method consistently surpasses state-of-the-art approaches in both quantitative metrics and qualitative perception.

This paper proposes a novel domain adaptation network to improve model accuracy in plant disease detection by leveraging structured image and text graphs. The proposed methodology constructs image graphs from segmented regions, including background, healthy, and disease regions. Class-specific text graphs are generated from prompts containing distinct color and texture attributes for each class. By aligning the structural information between image and text graphs, the network facilitates the effective transfer of class distinction capabilities from a source domain rich in data to a target domain with limited data, thereby addressing a key challenge of domain shift and data scarcity in agriculture. This approach incorporates text-image graph loss and image graph loss to enable comprehensive cross-domain class alignment, enhancing class distinction ability and ensuring consistent performance across diverse environmental conditions. Experimental evaluations across multiple source–target dataset pairs demonstrate improved class separation and stable mean IoU performance compared to existing approaches, validating the effectiveness of the proposed model in real-world agricultural scenarios. This method offers a promising pathway to enhance model adaptability and accuracy in disease detection, especially in complex and various agricultural environments.

This paper presents an empirical investigation into illuminant estimation using multi-spectral images. Our study emphasizes two key contributions: (1) the utilization of the estimated multi-spectral images and (2) the incorporation of a hierarchical structure. Firstly, exploiting multi-spectral images proves to have a positive influence on illuminant estimation, particularly in scenarios characterized by monochromatic images where conventional color constancy methods face challenges. Our experimental results vividly illustrate the effectiveness of leveraging spectral information in enhancing illuminant estimation. Secondly, the adoption of a hierarchical structure stems from the need for spatial invariance in the task of estimating a global illuminant. To further enhance the performance of the hierarchical structure, we employ a contrastive loss applied to different scaled outputs. This approach demonstrates remarkable effectiveness on our custom dataset, showcasing superior performance compared to the existing methods. In addition, we extend the evaluation to the widely recognized NUS-8 dataset, where the proposed method showcases a notable 26.7% relative improvement over the previous state-of-the-art methods.

Near-infrared (NIR) imaging captures more details and textures with less noise in low-light environments compared to RGB, making it widely used in such scenarios. However, the lack of color in NIR poses challenges for human cognition and computer vision, necessitating its colorization. We propose a multi-band NIR imaging approach with dual-teacher knowledge distillation to better estimate original color and structure. The dual-teacher network, with color- and structure-teacher, separately instructs the student network on color and structural qualities. To fuse these features, the Color Guided Structure (CGS) and the Color Embedding (CE) modules are applied. The CGS module enhances correlation by synchronizing color and structure under the guidance of the color feature, while the CE module effectively fuses them. Our model retains color consistency and detailed structure information of objects. The source code and datasets will be available at https://github.com/ymin2570/Multi-Band-NIR-colorization/.

Low-light images often lack visibility and information, and traditional methods of adjusting camera sensitivity and exposure time can result in visual quality degradation. In this paper, we propose a low-light enhancement method that utilizes two novel approaches to address these issues. The first approach involves using multi-band NIR (Near Infra-red) to preserve structural components, while a transformer-based cross-attention module efficiently calculates the correlation between NIR and RGB for effective fusion. The second approach involves implementing dual-teacher knowledge distillation, where normal- and mid-light teacher networks transfer low-light enhancement knowledge to the student. Our proposed method produces better color and detail restoration results than existing methods, particularly in ultra low-light environments. We also provide our own datasets for two different low-light conditions, enabling wide evaluations and ablation studies.

This paper proposes to utilize deep spatio-spectral features for color constancy, while most conventional methods focus on only spatial information of RGB images. We propose a novel color constancy method that exploits multi-band near-infrared (NIR). The proposed illuminant estimation network comprises two subnetworks, each tasked with learning different aspects: one focuses on capturing the local illuminant based on the input RGB and NIR images, while the other is dedicated to estimating confidence levels using both RGB and NIR inputs. We leverage a cross-attention mechanism to effectively capture the correlation between RGB and NIR modalities. Additionally, our method capitalizes on the relationship between RGB and NIR by incorporating a NIR colorization network into the primary illuminant estimation framework. Due to the absence of a public dataset for color constancy with multi-band NIR, we constructed a dataset comprising 604 images using a hyperspectral camera. The experimental results demonstrate superior performance in illuminant estimation when utilizing multi-band NIR compared to conventional RGB and single-band NIR approaches. Furthermore, employing a cross-attention mechanism to consider the correlation between RGB and NIR reveals the estimation of a sophisticated confidence map that aids in illuminant estimation, surpassing other existing fusion methods. The feature transfer from an additional task network can further enhance the performance of illuminant estimation.

We aim to contribute to deep learning based smart agriculture through semantic segmentation on crop images from real field environment. The key objective is the precise detection of diseases to facilitate the automation of agricultural management. The most significant issue is that the disease regions, serving as Regions of Interest (RoI), are small, making accurate prediction challenging. To address this issue, we propose a new framework of RoI-Attention Network (RA-Net) which additionally utilizes an RoI-attentive image that includes only regions predicted as disease and their surroundings from the input image. Using the RoI-attentive image, RA-Net enhances the representation power for disease regions by guiding the network to re-focus on RoI-associated context based on the initial prediction from the input. Using the proposed RoI-Attention stage, the coarse predictions of disease regions in crop images can be enhanced by incorporating additional sequential RoI-Attention and fusion stages. We have experimentally demonstrated the effectiveness of the proposed RA-Net in predicting small disease regions.

This paper proposes a network for plant and disease segmentation through semi-supervised learning in order to enhance agricultural production. Because of the hardness to get precisely labeled data, we use unlabeled data with pseudo labels. Furthermore, we employ a teacher-student network framework, where the teacher network imparts knowledge from labeled data to the student network. This boosts the segmentation precision of the student network, which is then exclusively trained on unlabeled data. We introduce a novel Exponential Moving Average (EMA) technique for teacher parameter updates, enhancing segmentation accuracy. Experimental results show that better segment performance can be achieved with the proposed network.

This paper proposes an image formation model that jointly combines dichromatic and intrinsic image decomposition models. The two decomposition models analyze image formation process from a different perspective, and they can be combined synergistically. It is confirmed that the proposed method performs better than the individual decomposition. The joint estimation and the study of the decomposition order (‘intrinsic + dichromatic’ or ‘dichromatic + intrinsic’) are the first attempt to the best of our knowledge. It was confirmed that the proposed ‘intrinsic + dichromatic’ is more optimal through experimental evaluations. We also exploit the temporal property of AC light sources, which can further improve the decomposition performance. The experimental results show that the proposed model can make an accurate image decomposition and achieve a remarkable color constancy performance.