Earticle

CNN Convolutional Neural Networks)은 영상 분류, 인식 및 검색 작업에 대한 유망한 결과를 보여주었다. 이 러한 관점에서, 스포츠 비디오 분류는 CNN이 덜 탐구된 능동적이고 도전적인 영역으로 남아 있다. 이에 우리는 새 로운 데이터 세트를 생성하여 스포츠 비디오 분류에 대한 CNN의 경험적 평가를 광범위하게 제공한다. 본 논문에서 는 MobileNetV2 (MbNetV2)네트워크를 이용한 CNN 기반 방법과 스포츠 비디오 분류를 위한 롤링 예측 평균 방법을 제안한다. 제안된 방법은 미세조정된 MbNetV2를 사용하여 비디오의 각 프레임을 분류하고 그 예측을 목록 에 저장한다. 롤링 예측 평균에서 마지막 "K" 예측의 평균이 계산되고 프레임에서 가장 높은 확률 레이블이 할당된 다. 우리는 제안한 방법이 스포츠 데이터 세트에서 97.9%의 최고 정확도를 달성한다는 것을 실험적으로 증명한다.

Convolutional Neural Networks(CNNs) have shown encouraging results for image classification, recognition, and retrieval tasks. In this perspective, the sport videos classification remains an active and challenging area where CNNs are less explored. Encouraged by this, we extensively provide an empirical evaluation of CNNs on sport videos classification by creating a new dataset. In this paper, we propose a CNN based method that uses MobileNetV2(MbNetV2) network and a rolling prediction average method for sport videos classification. The proposed method uses fine-tuned MbNetV2 to classify each frame in the video and stores its prediction in a list. In rolling predition average the mean of last "K" predictions is calculated and assigned the highest probability label to the frame. We experimentally prove that our proposed method achieves the best accuracy of 97.9% on our sport dataset.

Head pose estimation from a monocular image is crucial for applications in computer vision, AR/VR, and human?computer interaction. However, it remains challenging due to occlusions, lighting variations, and limited data. Landmark-based methods often suffer from localization errors, while landmark-free models tend to be complex and computationally expensive. To address these issues, we propose a lightweight, landmark-free CNN regressor guided by anthropometric facial measures. The model comprises two components: an Anthropometric Facial Measure Regressor (AFMR) that estimates a 4D vector of key facial segment lengths, and a CNN-based module that generates five uncertainty-based facial heatmaps. Evaluations on the BIWI and AFLW datasets show that our method outperforms state-of-the-art approaches, reducing localization error by 0.13° and 0.67°, respectively, while achieving faster convergence, lower parameter count, and real-time suitability.

This paper presents BlackNet, a vision-based black ice detection system designed for real-time vehicular safety. Unlike traditional methods that require expensive environmental sensors, BlackNet leverages existing onboard surround-view cameras. The proposed architecture integrates ResNet-style residual connections into a lightweight MobileNetV2 backbone to optimize feature extraction for subtle road surface variations. To enhance visibility in low-light and high-glare conditions, Contrast Limited Adaptive Histogram Equalization (CLAHE) is utilized for image preprocessing. The model was trained and validated on a comprehensive dataset of 15,200 images, achieving an accuracy of 92.4%. We propose a cloud-assisted deployment framework where inference is performed remotely in cloud, overcoming the computational constraints of edge devices. This approach offers a scalable, hardware-efficient solution for autonomous and connected vehicle safety.

In orthogonal frequency division multiplexing (OFDM) systems, the existing timing synchronization (TS) methods are challenged by high computational complexity, large processing delay, and performance degradation in wireless scenarios. To alleviate these issues, a lightweight cascaded one-dimensional convolutional neural network (1-D CNN)-based TS scheme is proposed in this paper, which is developed from the joint perspective of considering cascaded mode and network lightweight. Specifically, we meticulously design two lightweight subnetworks. The first one is implemented with only one 1-D CNN layer due to the coarse timing offset task, which aims to reduce the search range of TS for the following subnetwork. Based on the narrowed search range of TS, the second subnetwork is also light-weighted with a two-layer 1-D CNN, which refines the estimation of timing offset. By jointly considering the cascaded mode and the lightweight 1-D CNN, the proposed scheme improves the TS correctness with reduced computational complexity and processing delay compared with the existing works. Simulation results validate the effectiveness and robustness of the proposed TS scheme given variant parameters.

This paper investigates deep learning-based SNR estimation for OFDM systems. A lightweight ResNet-inspired model is applied to estimate SNR under AWGN, Rayleigh, and Rician channels. Specifically, our model consists of two residual blocks to ensure a lightweight design. The dataset includes wide SNR ranges with realistic impairments such as fading and frequency offsets. Performance is evaluated using mean square error (MSE) and mean absolute error (MAE). Results show stable estimation across all channels with low error values in the low SNR regions.

Nowadays, due to natural disasters the world is facing huge challenges such as economical, climatic, and losses a lot of precious human life. The traditional emergency response and rescue teams are physically visit different affected areas for inspection and save human lives. In this manual monitoring system created various problems such as human resources, time-consuming, and in real-time unable to accurately analyze the nature of the disaster. Therefore, there is an urgent need for an automatic real-time system to intelligently identified different disaster scenes and analyze the affected areas for quick response. Therefore, in this paper, an Unmanned Aerial Vehicles (UAVs) inspired framework is proposed for disaster scenes classification using a lightweight Convolution Neural Network (CNN). To validate the strength of the proposed framework a comparative analysis is conducted to show its superiority against different state-of-the-art models in terms of computational complexity and performance.

본 연구는 온라인 리뷰의 유용성을 예측하기 위해 텍스트 의미와 구조 정보를 통합한 BERT-CNN 기반 딥러닝 모델을 제안한다. 자원 효율성이 높은 사전학습 언어모델인 DistilBERT, ALBERT, TibyBERT를 활용하여 문맥 임베딩을 생성하고, 이를 CNN 구조와 결합해 문장 내 지역적 의미 패턴을 효과적으로 포착하였다. 또한 텍스트에서 추출한 복잡성, 가독성, 감정/감성 등의 설계 기반 특성과 사용자 및 비즈니스 메타데이터를 통합하여 임베딩만으로는 포착하기 어려운 도메인 특성을 보완하였다. 제안한 CNN 기반 하이브리드 모델은 사전학습 언어모델 단독 구조보다 전반적으로 높은 예측 성능을 보였다. 특히 TinyBERT-CNN 모델은 가장 높은 성과를 나타내어 경량 모델의 효율성과 실용성을 입증하였다. DistilBERT 및 ALBERT 기반 모델 역시 언어적 특성과 메타데이터의 결합을 통해 단일 모델 대비 정확도와 정밀도가 향상되었으며, 이는 수치화된 언어 특성과 메타데이터가 문맥 임베딩 기반 모델을 보완할 수 있음을 보여준다.

본 논문은 재난 환경에 적합한 경량화된 UWB 측위에 대한 연구를 소개한다. UWB는 디바이스간 정 밀한 거리 측정 기능을 제공하여 고정밀 측위에 활용된다. 하지만 재난 환경과 같이 무선 채널이 복잡 한 경우, UWB 측위 성능이 급격히 하락하는 문제가 발생한다. 최근 UWB 신호에 딥러닝 기술을 이용 하여 성능 하락 문제를 해결하는 연구들이 제안된다. 하지만 딥러닝 기술은 고성능 컴퓨팅 자원을 요 구하여 자원 제약이 있는 재난 상황에서 활용되기 어렵다. 본 논문에서는 XAI 기술을 적용하여 딥러 닝 기술의 높은 정확도를 유지하면서 동시에 연산 복잡도를 낮추기 위한 XLNet을 제안한다.

Image processing is one of the major techniques that are used for computer vision. Nowadays, researchers are using machine learning and deep learning for the aforementioned task. In recent years, digit recognition tasks, i.e., automatic meter recognition approach using electric or water meters, have been studied several times. However, two major issues arise when we talk about previous studies: first, the use of the deep learning technique, which includes a large number of parameters that increase the computational cost and consume more power; and second, recent studies are limited to the detection of digits and not storing or providing detected digits to a database or mobile applications. This paper proposes a system that can detect the digital number of meter readings using a lightweight deep neural network (DNN) for low power consumption and send those digits to an Android mobile application in real-time to store them and make life easy. The proposed lightweight DNN is computationally inexpensive and exhibits accuracy similar to those of conventional DNNs.

This study estimates heart rate (HR) from piezo-based ballistocardiography (BCG) embedded in a bed topper(or mattress topper) using a five-channel acquisition and a multi-scale STFT spectrogram(windows 128,256,512) representation (final tensor 15×64×45) with a lightweight CNN regressor, and deploys it on a Hailo-8 NPU to jointly validate latency and accuracy. Preprocessing uses 100 Hz sampling and a 0.83~2.0Hz zero-phase Butterworth band-pass filter, labels come from synchronous PPG inter-peak intervals and are z-normalized (𝜇=75, 𝜎=15). The PyTorch model is converted via ONNX→HAR, quantized with a calibration set, compiled to HEF, and executed on device. Data splits (80%/10%/10%) and metrics (MAE, RMSE, correlation, mean latency) follow internal scripts. Increasing the channel count reduces MAE by ~27.3% from one to five channels and yields r=0.8780, combining multi-scale STFTs reduces MAE by ~23.7% over single-scale and achieves r=0.9023. Versus an FFT peak-picking baseline, MAE drops by ~66.6% and correlation rises from r=0.4934 to 0.9023. INT8 deployment attains MAE 1.84 BPM (+7.6% vs float) with 2.91 ms inference latency. Overall, multi-channel/multi-scale fusion with a compact CNN provides robust, real-time BCG-HR estimation on edge hardware.

The casting manufacturing process of aluminum automotive wheels often involves processing various wheel models during stages such as flow forming, machining, packaging, and delivery. Traditionally, separate equipment or production lines were required for each model, which led to higher facility investment costs and increased labor costs for classification. However, the implementation of machine learning-based model classification technology has made it possible to automatically and accurately distinguish between different wheel models, resulting in significant cost savings and enhanced production efficiency. Additionally, this approach helps prevent product mix-ups during the final inspection process and allows for the quick and precise identification of wheel models during packaging and delivery, reducing shipping errors and improving customer satisfaction. Despite these benefits, the high cost of machine learning equipment presents a challenge for small and medium-sized enterprises(SMEs) to adopt such technologies. Therefore, this paper analyzes the characteristics of existing machine learning architectures applicable to the automotive wheel manufacturing process and proposes a custom CNN(Convolutional Neural Network) that can be used efficiently and cost-effectively.

위성 영상 촬영 후 지상국에 전송된 영상을 이용하여 최종 위성 영상을 획득하기 위해 많은 영상 전/후 처리 과정이 수반된다. 전/후처리 과정 중 레벨 1R 영상에서 레벨 1G 영상으로 변환 시 기하 보정은 필수적으로 요구된다. 기하 보정 알고리즘에서는 보간 기법을 필연적으로 사용하게 되며, 보간 기법의 정확도에 따라서 레벨 1G 영상의 품질이 결정된다. 또한, 레벨 프로세서에서 수행되는 보간 알고리즘의 고속화 역시 매우 중요하다. 본 논문에서는 레벨 1R에서 레벨 1G로 변환 시 기하 보정에 필요한 경량화된 심층 컨볼루션 신경망 기반 보간 기법에 대해 제안하였다. 제안한 기법은 위성 영상의 해상도를 2배 향상하며, 빠른 처리 속도를 위해 경량화된 심층 컨볼루션 신경망으로 딥러닝 네트워크를 구성하였다. 또한, panchromatic (PAN) 밴드 정보를 활용하여 multispectral (MS) 밴드의 영상 품질 개선이 가능한 피처 맵 융합 방법을 제안하였다. 제안된 보간 기술을 통해 획득한 영상은 기존의 딥러닝 기반 보간 기법에 비해 정량적인 peak signal-to-noise ratio (PSNR) 지표에서 PAN 영상은 약 0.4 dB, MS 영상은 약 4.9 dB 개선된 결과를 보여주었으며, PAN 영상 크기 기준 36,500×36,500 입력 영상의 해상도를 2배 향상된 영상 획득 시 기존 딥러닝 기반 보간 기법 대비 처리 속도가 약 1.6배 향상됨을 확인하였다.

In order to obtain satellite image products using the image transmitted to the ground station after capturing the satellite images, many image pre/post-processing steps are involved. During the pre/post-processing, when converting from level 1R images to level 1G images, geometric correction is essential. An interpolation method necessary for geometric correction is inevitably used, and the quality of the level 1G images is determined according to the accuracy of the interpolation method. Also, it is crucial to speed up the interpolation algorithm by the level processor. In this paper, we proposed a lightweight CNN-based interpolation method required for geometric correction when converting from level 1R to level 1G. The proposed method doubles the resolution of satellite images and constructs a deep learning network with a lightweight deep convolutional neural network for fast processing speed. In addition, a feature map fusion method capable of improving the image quality of multispectral (MS) bands using panchromatic (PAN) band information was proposed. The images obtained through the proposed interpolation method improved by about 0.4 dB for the PAN image and about 4.9 dB for the MS image in the quantitative peak signal-to-noise ratio (PSNR) index compared to the existing deep learning-based interpolation methods. In addition, it was confirmed that the time required to acquire an image that is twice the resolution of the 36,500×36,500 input image based on the PAN image size is improved by about 1.6 times compared to the existing deep learning-based interpolation method.

엣지 디바이스에서 딥러닝을 활용하기 위하여 CNN 경량화 연구들이 진행되고 있다. 경량 CNN 은 대부분 고정 소수점을 사용하며, 계층에 따라 정밀도는 달라진다. 본 논문에서는 경량 CNN 을 지원하기 위하여, 사용 계층에 따라 정밀도를 선택할 수 있는 가변 정밀도 곱셈기를 제안한다. 제안하는 가변 정밀도 곱셈기는 낮은 정밀도 곱셈기를 병합하는 구조로, 정밀도가 낮을 때는 병렬 처리를 통해 효율을 높인다. 제안하는 곱셈기를 Verilog HDL로 설계하고 ModelSim 에서 동작을 확인하였다. 설계된 곱셈기는 계층별로 정밀도가 다른 CNN 가속기에서 효율적으로 적용될 것으로 기대된다.

객체 검출 알고리즘은 자율주행 시스템 구현을 위한 핵심 요소이다. 최근 심층 합성곱 신경망 (Deep Convolutional Neural Network) 기반의 영상 인식 기술이 발전함에 따라 심층 학습을 이용한 객체 검출 관련 연구들이 활발히 진행되고 있다. 본 논문에서는 객체 검출에 가장 널리 사용되고 있는 Mask R-CNN의 경량화 모델을 제안하여 도로 내 다양한 객체들의 위치와 형태를 효율적으로 예측하는 방법을 제안한다. 또한, 주의 모듈(Attention Module)을 Mask R-CNN 내 각각 다른 역할을 수행하는 신경망 계층에 적용함으로써 특징 지도를 적응적으로 재교정(Re-calibration)하여 검출 성능을 향상시킨다. 실제 주행 영상에 대한 다양한 실험 결과를 통해 제안하는 방법이 기존 방법 대비 크게 감소된 신경망 매개변수만을 이용하여 고성능 검출 성능을 유지함을 보인다.

Object detection plays a crucial role in a self-driving system. With the advances of image recognition based on deep convolutional neural networks, researches on object detection have been actively explored. In this paper, we proposed a lightweight model of the mask R-CNN, which has been most widely used for object detection, to efficiently predict location and shape of various objects on the road environment. Furthermore, feature maps are adaptively re-calibrated to improve the detection performance by applying an attention module to the neural network layer that plays different roles within the mask R-CNN. Various experimental results for real driving scenes demonstrate that the proposed method is able to maintain the high detection performance with significantly reduced network parameters.