Asus (Taiwan)

In the realm of machine-learning (ML)-based electronic design automation (EDA), several factors contribute to inefficiency, posing various challenges. Initially, the lack of flexibility in input structures hinders the sharing of information across different circuit topologies. In addition, substantial costs are incurred in terms of simulation run times during the data generation process due to the necessity of creating a large training dataset for each circuit topology. To this effect, in this article, we address the dual problem of how to: 1) develop a general unified surrogate model that can handle a variety of circuit topologies and 2) employ previously trained models and adapt them to new models. We provide a formulation for transforming 3-D electromagnetic (EM) circuits into versatile circuit graphs, for a variety of topologies, imbued with structural information. The absence of such frameworks represents a gap in ML-based EDA, which we fill by providing a set of building blocks to achieve significant improvements in modeling tasks. Finally, we present a versatile forward modeling framework that allows one to quickly obtain the output response given a set of design parameters. We achieve the overarching goal of reducing the resources needed to create an ML model library for signal integrity (SI) applications in microelectronics packaging.

Power distribution networks (PDNs) are designed to deliver sufficient and stable power to printed circuit board (PCB) components. With increasing current/voltage demands on modern PCBs, power rails in PDN require more space for metals. Existing design techniques often adopt greedy heuristics to optimize the metal shapes of power rails. However, they consider little spatial competition among different power rails, so routing spaces may be distributed unevenly, causing redundant metal usage, current/voltage violations, and even routing failures in high-current designs. To remedy this problem, we propose a multi-layered PDN design flow to distribute rail spaces according to current/voltage demands appropriately. To our knowledge, this is the first work to consider the spatial competition issue in PCB power rail routing. Specifically, we generate a multi-layered topology and minimize the metal area and the via usage with mathematical programming. To resolve the overlap regions by rail detouring, we also propose a resistance-aware routing algorithm. Experimental results show that our work significantly outperforms a baseline algorithm extended from the state-of-the-art rail router in the metal area and runtime, achieving respective reductions of 55% and 71%, without any current/voltage violations.

This study addresses the rising concern of radio frequency interference (RFI) issue in high-speed connectors, particularly focus on comparing the radiation noise generated from differential-mode (DM) signals and common-mode (CM) noise. While prior research has concentrated on CM noises and their suppression methods due to their strong radiation efficiency, DM noises, despite possessing weaker radiation efficiency, have received less attention in past research. However, it is worth noting that DM signals possess at least 20 dB higher energy than CM signals in digital systems. To compare the radiation caused by CM noise and DM signals, a 3D model with USB 3.0 high-speed connector system is constructed and validated through full-wave simulations and near-field measurements. The simulated results reveal that radiation caused by DM signals exhibit comparable significance to that caused by CM noises in digital systems. This finding underscores the importance of considering DM signals and CM noises in digital systems and provides new insights into mitigating interference in electronic devices.

Current augmented reality (AR) waveguide display technology faces a challenging trade-off between essential factors such as eyebox, efficiency, and image uniformity. Surface relief grating has been a commonly used exit pupil expander (EPE) in waveguides, but it exhibits limitations in efficiency and uniformity. To overcome these shortcomings, polarized volume grating (PVG) with ultra-wide angular bandwidth and high efficiency has been employed as an EPE to enhance efficiency and uniformity. However, the challenge arises when attempting to achieve high uniformity within the eyebox using nematic liquid crystal for controlling each expansion, as the polarization changes of diffracted light contribute to varying diffraction efficiencies, which affects the maintenance of uniformity across the entire field of view. This article introduces an innovative approach by incorporating high-low cover-layers on top of PVG to effectively control the path difference between light rays at various angles. This design modification mitigates unnecessary out-coupling and ensures a remarkably high level of uniformity in the entire AR system, providing a more efficient and uniform expanded eyebox.

Single image haze removal is crucial in computer vision. In open literatures, two kinds of dehazing strategies (prior-based and learning-based methods) have been developed. However, they have a trade-off between detail preservation and the image quality. Prior-based methods reconstruct the detail well but have lower image quality while learning-based methods achieve better recovered quality but lose the detail. In this paper, to mitigate this dilemma, a hierarchical architecture using the discrete wavelet transform (DWT) is proposed. It divides the dehazing problem into two parts: detail and background reconstruction. Based on investigating how haze affects the image in the wavelet domain, two networks for detail and background reconstruction are proposed. To avoid color distortion and the detail loss, the anti-vanish wavelet loss and the bound penalty are proposed. The multi-level wavelet component discriminator is proposed for further improvement. Experiments show that the proposed network can achieve superior performance in all metrics.

Image relighting aims to recalibrate the illumination set?ting in an image. In this paper, we propose a deep learning?based method called multi-modal bifurcated network (MB?Net) for depth guided image relighting. That is, given an image and the corresponding depth maps, a new image with the given illuminant angle and color temperature is gener?ated by our network. This model extracts the image andthe depth features by the bifurcated network in the encoder. To use the two features effectively, we adopt the dynamic dilated pyramid modules in the decoder. Moreover, to in?crease the variety of training data, we propose a novel data process pipeline to increase the number of the training data. Experiments conducted on the VIDIT dataset show that the proposed solution obtains the 1st place in terms of SSIM and PMS in the NTIRE 2021 Depth Guide One-to-one Re?lighting Challenge.

Detecting human-object interactions (HOI) is an important step toward a comprehensive visual understanding of machines. While detecting non-temporal HOIs (e.g., sitting on a chair) from static images is feasible, it is unlikely even for humans to guess temporal-related HOIs (e.g., opening/closing a door) from a single video frame, where the neighboring frames play an essential role. However, conventional HOI methods operating on only static images have been used to predict temporal-related interactions, which is essentially guessing without temporal contexts and may lead to sub-optimal performance. In this paper, we bridge this gap by detecting video-based HOIs with explicit temporal information. We first show that a naive temporal-aware variant of a common action detection baseline does not work on video-based HOIs due to a feature-inconsistency issue. We then propose a simple yet effective architecture named Spatial-Temporal HOI Detection (ST-HOI) utilizing temporal information such as human and object trajectories, correctly-localized visual features, and spatial-temporal masking pose features. We construct a new video HOI benchmark dubbed VidHOI where our proposed approach serves as a solid baseline.

Underwater image enhancement is an important low-level computer vision task for autonomous underwater vehicles and remotely operated vehicles to explore and understand the underwater environments. Recently, deep convolutional neural networks (CNNs) have been successfully used in many computer vision problems, and so does underwater image enhancement. There are many deep-learning-based methods with impressive performance for underwater image enhancement, but their memory and model parameter costs are hindrances in practical application. To address this issue, we propose a lightweight adaptive feature fusion network (LAFFNet). The model is the encoder-decoder model with multiple adaptive feature fusion (AAF) modules. AAF subsumes multiple branches with different kernel sizes to generate multi-scale feature maps. Furthermore, channel attention is used to merge these feature maps adaptively. Our method reduces the number of parameters from 2.5M to 0.15M (around 94% reduction) but outperforms state-of-the-art algorithms by extensive experiments. Furthermore, we demonstrate our LAFFNet effectively improves high-level vision tasks like salience object detection and single image depth estimation.

The selection of decoupling capacitors (decap) is a critical but tedious process in power distribution network (PDN) design. In this paper, an improved decap-selection algorithm based on deep reinforcement learning (DRL), which seeks the minimum number of decaps through a self-exploration training to satisfy a given target impedance, is presented. Compared with the previous relevant work: the calculation speed of PDN impedance is significantly increased by adopting an impedance matrix reduction method; also, the enhanced algorithm performs a better convergence by utilizing the techniques of double Q-learning and prioritized experience replay; furthermore, a well-designed reward is proposed to facilitate long-term convergence when more decaps are required. The proposed algorithm demonstrates the feasibility of achieving decent performance using DRL with pre-trained knowledge for more complicated engineering tasks in the future.

Person re-identification (re-ID) aims at recognizing the same person from images taken across different cameras. To address this challenging task, existing re-ID models typically rely on a large amount of labeled training data, which is not practical for real-world applications. To alleviate this limitation, researchers now targets at cross-dataset re-ID which focuses on generalizing the discriminative ability to the unlabeled target domain when given a labeled source domain dataset. To achieve this goal, our proposed Pose Disentanglement and Adaptation Network (PDA-Net) aims at learning deep image representation with pose and domain information properly disentangled. With the learned cross-domain pose invariant feature space, our proposed PDA-Net is able to perform pose disentanglement across domains without supervision in identities, and the resulting features can be applied to cross-dataset re-ID. Both of our qualitative and quantitative results on two benchmark datasets confirm the effectiveness of our approach and its superiority over the state-of-the-art cross-dataset Re-ID approaches.