Designing vehicle exteriors requires repeated refinement of key proportions and viewpoints, a process traditionally reliant on manual sketching, which is often time-consuming and inefficient in early concept stages. To accelerate the design process, we are exploring the potential of utilizing AI for ideation in these early stages. However, it remains a challenging task to control proportions and maintain a fixed perspective when generating images using AI. To address these limitations, we present RatioMorph, a controllable image generation system that enables manipulation of vehicle proportions and viewpoints when generating images by AI. RatioMorph comprises two core modules. Car2BoxNet is a depth estimation model that transforms real photographs into structured box-style depth maps that capture the geometric layout of the vehicle. Box2CarNet is a diffusion-based image generator fine-tuned to produce vehicle designs that adhere to the provided geometric conditions. Both Car2BoxNet and Box2CarNet are trained on a synthetic dataset curated through automated filtering based on geometric alignment and visual quality. Evaluated within a production-adjacent automotive design workflow, RatioMorph significantly reduced early-stage design iteration time and enabled exploratory workflows that were difficult with previous AI workflows. This work introduces a domain-specific, controllable diffusion-based generation system tailored for automotive design, enabling manipulation of vehicle viewpoint and proportion. It demonstrates strong potential to accelerate early-stage workflows and outlines a path toward industrial deployment, with phased integration into production environments currently underway.

Rainfall rate (RR) retrieval methods utilizing visible (VIS), near-infrared (NIR), and infrared (IR) bands from geostationary (GEO) meteorological satellites often achieve higher accuracy compared to methods relying solely on IR bands. However, these methods face limitations for nighttime applications. This study aims to introduce an artificial intelligence (AI)-based approach for nighttime RR retrieval using actual IR observations and virtual nighttime VIS (0.47, 0.51, and 0.86 μm) and NIR (1.37 and 1.61 μm) bands data generated through an adversarial data-to-data (D2D) translation method. The study focuses on the Korean Peninsula and employs paired observational datasets from the Advanced Meteorological Imager (AMI) on Geo-KOMPSAT-2A (GK2A) and hybrid surface rainfall data from ground weather radars as input and target domains for the D2D model. The study generates virtual D2D-Generated (DG) VIS and NIR band data and evaluates DG-RR using only IR bands during nighttime. Results indicate that DG-RR using virtual AMI VIS, NIR, and real AMI IR bands demonstrates superior quantitative and qualitative statistical scores compared to DG-RR based solely on the AMI IR band data. This improvement is observed over traditional algorithm-based GK2A AMI RR and precipitation estimation from remotely sensed information using artificial neural networks. Thus, this study underscores the potential of utilizing virtual AI-based VIS and NIR band data, resembling actual observations, in RR estimation and various satellite remote sensing applications.

When training early-stage deep neural networks (DNNs), generating intermediate features via convolution or linear layers occupied most of the execution time. Accordingly, extensive research has been done to reduce the computational burden of the convolution or linear layers. In recent mobile-friendly DNNs, however, the relative number of operations involved in processing these layers has significantly reduced. As a result, the proportion of the execution time of other layers, such as batch normalization layers, has increased. Thus, in this work, we conduct a detailed analysis of the batch normalization layer to efficiently reduce the runtime overhead in the batch normalization process. Backed up by the thorough analysis, we present an extremely efficient batch normalization, named LightNorm, and its associated hardware module. In more detail, we fuse three approximation techniques that are i) low bit-precision, ii) range batch normalization, and iii) block floating point. All these approximate techniques are carefully utilized not only to maintain the statistics of intermediate feature maps, but also to minimize the off-chip memory accesses. By using the proposed LightNorm hardware, we can achieve significant area and energy savings during the DNN training without hurting the training accuracy. This makes the proposed hardware a great candidate for the on-device training.

Recent advances in deep learning have made it possible to implement artificial intelligence in mobile devices. Many studies have put a lot of effort into developing lightweight deep learning models optimized for mobile devices. To overcome the performance limitations of manually designed deep learning models, an automated search algorithm, called neural architecture search ( NAS ), has been proposed. However, studies on the effect of hardware architecture of the mobile device on the performance of NAS have been less explored. In this article, we show the importance of optimizing a hardware architecture, namely, NPU dataflow, when searching for a more accurate yet fast deep learning model. To do so, we first implement an optimization framework, named FlowOptimizer, for generating a best possible NPU dataflow for a given deep learning operator. Then, we utilize this framework during the latency-aware NAS to find the model with the highest accuracy satisfying the latency constraint. As a result, we show that the searched model with FlowOptimizer outperforms the performance by 87.1% and 92.3% on average compared to the searched model with NVDLA and Eyeriss, respectively, with better accuracy on a proxy dataset. We also show that the searched model can be transferred to a larger model to classify a more complex image dataset, i.e., ImageNet, achieving 0.2%/5.4% higher Top-1/Top-5 accuracy compared to MobileNetV2-1.0 with 3.6 \( \times \) lower latency.