JLJunmin Lee

Junmin Lee

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Junmin Lee has a total of 8 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Real-Time Monte Carlo Denoising With Adaptive Fusion Network

    by Junmin Lee, Seunghyun Lee, Min Yoon, Byung Cheol Song

    Real-time Monte Carlo denoising aims to denoise a 1spp-rendered image with a limited time budget. Many latest techniques for real-time Monte Carlo denoising utilize temporal accumulation (TA) as a pre-processing to improve the temporal stability of successive frames and increase the effective spp. However, existing techniques using TA used to suffer from significant performance degradation when TA does not work well. In addition, they have the disadvantage of deteriorating performance in dynamic scenes because pixel information of the current frame cannot be sufficiently utilized due to the pixel averaging effect between temporally adjacent frames. To solve this problem, this paper proposes a framework that utilizes both 1spp images and temporally accumulated 1spp (TA-1spp) images. First, the multi-scale kernel prediction module estimates kernel maps for filtering 1spp images and TA-1spp images, respectively. Then, the filtered images are properly fused so that the two advantages of 1spp and TA-1spp images can create synergy. Also, the remaining noise is removed through the refinement module and fine details are reconstructed to improve the model flexibility, beyond using only the kernel prediction module. As a result, we achieve better quantitative and qualitative performance at 39% faster than state-of-the-art (SOTA) real-time Monte Carlo denoisers.


  • Open Access

    Display Visibility Improvement Through Content and Ambient Light-Adaptive Image Enhancement

    by Junmin Lee, Heejin Lee, Seunghyun Lee, Junho Heo, Jiwon Lee, Byung Cheol Song

    An image in a display device under strong illuminance can be perceived as darker than the original due to the nature of the human visual system (HVS). In order to alleviate this degradation in terms of software, existing schemes employ global luminance compensation or tone mapping. However, since such approaches focus on restoring luminance only, it has a fundamental drawback that chrominance cannot be sufficiently restored. Also, the previous approaches seldom provide acceptable visibility because it does not consider the content of an input image. Furthermore, because they focus mainly on global image quality, they may show unsatisfactory image quality for certain local areas. This paper introduces VisibilityNet, a neural network model designed to restore both chrominance and luminance. By leveraging VisibilityNet, we generate an optimally enhanced dataset tailored to the ambient light conditions. Furthermore, employing the generated dataset and a convolutional neural network (CNN), we estimate weighted piece-wise linear enhancement curves (WPLECs) that take into account both ambient light and image content. These WPLECs effectively enhance global contrast by addressing both luminance and chrominance aspects. Ultimately, through the utilization of a salient object detection algorithm that emulates the HVS, visibility enhancement is achieved not only for the overall region but also for visually salient areas. We verified the performance of the proposed method by comparing it with five existing approaches in terms of two quantitative metrics for a dataset we built ourselves. Experimental findings substantiate that the proposed method surpasses alternative approaches by significantly improving visibility.


  • Open Access

    Metric-Based Attention Feature Learning for Video Action Recognition

    by Dae Ha Kim, Fazliddin Anvarov, Jun Min Lee, Byung Cheol Song

    Conventional approaches for video action recognition were designed to learn feature maps using 3D convolutional neural networks (CNNs). For better action recognition, they trained the large-scale video datasets with the representation power of 3D CNN. However, action recognition is still a challenging task. Since the previous methods rarely distinguish human body from environment, they often overfit background scenes. Note that separating human body from background allows to learn distinct representations of human action. This paper proposes a novel attention module aiming at only action part(s), while neglecting non-action part(s) such as background. First, the attention module employs triplet loss to differentiate active features from non-active or less active features. Second, two attention modules based on spatial and channel domains are proposed to enhance the feature representation ability for action recognition. The spatial attention module is to learn spatial correlation of features, and the channel attention module is to learn channel correlation. Experimental results show that the proposed method achieves state-of-the-art performance of 41.41% and 55.21% on Diving48 and Something-V1 datasets, respectively. In addition, the proposed method provides competitive performance even on UCF101 and HMDB-51 datasets, i.e., 95.83% on UCF-101 and 74.33% on HMDB-51.