HyunWook Park

Hyunwook Park, Minsu Kim, Seongguk Kim et al.

In this article, for the first time, we propose a transformer network-based reinforcement learning (RL) method for power distribution network (PDN) optimization of high bandwidth memory (HBM). The proposed method can provide an optimal decoupling capacitor (decap) design to maximize the reduction of PDN self- and transfer impedance seen at multiple ports. An attention-based transformer network is implemented to directly parameterize decap optimization policy. The optimality performance is significantly improved since the attention mechanism has powerful expression to explore massive combinatorial space for decap assignments. Moreover, it can capture sequential relationships between the decap assignments. The computing time for optimization is dramatically reduced due to the reusable network on positions of probing ports and decap assignment candidates. This is because the transformer network has a context embedding process to capture meta-features including probing ports positions. In addition, the network is trained with randomly generated data sets. Therefore, without additional training, the trained network can solve new decap optimization problems. The computing time for training and data cost are critically decreased due to the scalability of the network. Thanks to its shared weight property, the network can adapt to a larger scale of problems without additional training. For verification, we compare the results with conventional genetic algorithm (GA), random search (RS), and all the previous RL-based methods. As a result, the proposed method outperforms in all the following aspects: optimality performance, computing time, and data efficiency.

Sunho Kim, Byungjai Kim, HyunWook Park

Medical image analysis using deep neural networks has been actively studied. Deep neural networks are trained by learning data. For accurate training of deep neural networks, the learning data should be sufficient, of good quality, and should have a generalized property. However, in medical images, it is difficult to acquire sufficient patient data because of the difficulty of patient recruitment, the burden of annotation of lesions by experts, and the invasion of patients' privacy. In comparison, the medical images of healthy volunteers can be easily acquired. Using healthy brain images, the proposed method synthesizes multi-contrast magnetic resonance images of brain tumors. Because tumors have complex features, the proposed method simplifies them into concentric circles that are easily controllable. Then it converts the concentric circles into various realistic shapes of tumors through deep neural networks. Because numerous healthy brain images are easily available, our method can synthesize a huge number of the brain tumor images with various concentric circles. We performed qualitative and quantitative analysis to assess the usefulness of augmented data from the proposed method. Intuitive and interesting experimental results are available online at https://github.com/KSH0660/BrainTumor

Giyong Choi, HyunWook Park

In contrast to single-image deblurring, video deblurring has the advantage that neighbor frames can be utilized to deblur a target frame. However, existing video deblurring algorithms often fail to properly employ the neighbor frames, resulting in sub-optimal performance. In this paper, we propose a deconvolution and aggregation network (DAN) for video deblurring that utilizes the information of neighbor frames well. In DAN, both deconvolution and aggregation strategies are achieved through three sub-networks: the preprocessing network (PPN) and the alignment-based deconvolution network (ABDN) for the deconvolution scheme; the frame aggregation network (FAN) for the aggregation scheme. In the deconvolution part, blurry inputs are first preprocessed by the PPN with non-local operations. Then, the output frames from the PPN are deblurred by the ABDN based on the frame alignment. In the FAN, these deblurred frames from the deconvolution part are combined into a latent frame according to reliability maps which infer pixel-wise sharpness. The proper combination of three sub-networks can achieve favorable performance on video deblurring by using the neighbor frames suitably. In experiments, the proposed DAN was demonstrated to be superior to existing state-of-the-art methods through both quantitative and qualitative evaluations on the public datasets.

Giyong Choi, HyunWook Park

Video quality is often severely degraded by multiple factors rather than a single factor. These low-quality videos can be restored to high-quality videos by sequentially performing appropriate video enhancement techniques. However, the sequential approach was inefficient and sub-optimal because most video enhancement approaches were designed without taking into account that multiple factors together degrade video quality. In this paper, we propose a new joint video enhancement method that mitigates multiple degradation factors simultaneously by resolving an integrated enhancement problem. Our proposed network, named DSFN, directly produces a high-resolution, high-frame-rate, and clear video from a low-resolution, low-frame-rate, and blurry video. In the DSFN, low-resolution and blurry input frames are enhanced by a joint deblurring and super-resolution (JDSR) module. Meanwhile, intermediate frames between input adjacent frames are interpolated by a triple-frame-based frame interpolation (TFBFI) module. The proper combination of the proposed modules of DSFN can achieve superior performance on the joint video enhancement task. Experimental results show that the proposed method outperforms other sequential state-of-the-art techniques on public datasets with a smaller network size and faster processing time.

Yongjin Lee, Seokjun Park, Beomgu Kang et al.

The paper proposes an image-guided depth completion method to estimate accurate dense depth maps with fast computation time. The proposed network has two-stage structure. The first stage predicts a first depth map. Then, the second stage further refines the first depth map using the confidence maps. The second stage consists of two layers, each of which focuses on different regions and generates a refined depth map and a confidence map. The final depth map is obtained by combining two depth maps from the second stage using the corresponding confidence maps. Compared with the top-ranked models on the KITTI depth completion online leaderboard, the proposed model shows much faster computation time and competitive performance.

Byungjai Kim, Kinam Kwon, Changheun Oh et al.

Anomaly detection in medical imaging is to distinguish the relevant biomarkers of diseases from those of normal tissues. Deep supervised learning methods have shown potentials in various detection tasks, but its performances would be limited in medical imaging fields where collecting annotated anomaly data is limited and labor-intensive. Therefore, unsupervised anomaly detection can be an effective tool for clinical practices, which uses only unlabeled normal images as training data. In this paper, we developed an unsupervised learning framework for pixel-wise anomaly detection in multi-contrast magnetic resonance imaging (MRI). The framework has two steps of feature generation and density estimation with Gaussian mixture model (GMM). A feature is derived through the learning of contrast-to-contrast translation that effectively captures the normal tissue characteristics in multi-contrast MRI. The feature is collaboratively used with another feature that is the low-dimensional representation of multi-contrast images. In density estimation using GMM, a simple but efficient way is introduced to handle the singularity problem which interrupts the joint learning process. The proposed method outperforms previous anomaly detection approaches. Quantitative and qualitative analyses demonstrate the effectiveness of the proposed method in anomaly detection for multi-contrast MRI.

Jee Won Kim, Kinam Kwon, Byungjai Kim et al.

An attention guided scheme for metal artifact correction in MRI using deep neural network is proposed in this paper. The inputs of the networks are two distorted images obtained with dual-polarity readout gradients. With MR image generation module and the additional data consistency loss to the previous work [1], the network is trained to estimate the frequency-shift map, off-resonance map, and attention map. The attention map helps to produce better distortion-corrected images by weighting on more relevant distortion-corrected images where two distortion-corrected images are produced with half of the frequency-shift maps. In this paper, we observed that in a real MRI environment, two distorted images obtained with opposite polarities of readout gradient showed artifacts in a different region. Therefore, we proved that using the attention map was important in that it reduced the residual ripple and pile-up artifacts near metallic implants.