Earticle

최근 클라우드 기반의 자원들은 다양한 장치를 통해 클라우드 서버로 저장하여 사용되고 있지만, 클라우드 서버로 전송되는 자원들은 클라우드에서 작업량을 최적화하기 위해 특정한 기능을 연결해주는 유연성 및 이동성에 제약이 있다. 본 논문에서는 속도, 성능, 안정성과 같은 요소를 기준으로 클라우드에서 작업량을 최적화할 수 있는 멀티 클라우드 기반 시스템 장애 최소화를 위한 분산 클라우드 자원 할당 기법을 제안한다. 제안 기법은 클라우드 자원의 특징을 계층별로 추출하여 특징들의 패턴을 합성 과정을 통해 클라우드 자원 할당을 동적으로 연계하여 최적화한다. 특히, 제안 기법은 멀티 클라우드 환경에서 클라우드 자원 할당을 최소화하기 위해서 동일 패턴의 클 라우드 자원을 블록체인으로 묶고, 합성 과정을 통해 클라우드 자원 크기를 줄임으로써 클라우드 자원을 일관성 있게 유지한다. 성능평가 결과, 제안 기법은 클라우드 자원 간 확률값에 따라 서로 연계하기 때문에 클라우드 자원 검증 시간이 평균 7.42% 향상되었고, 블록체인의 크기가 증가할수록 검증 지연 시간은 평균 8.81% 낮게 나타났다. 또한, 블록체인 그룹 크기에 따른 클라우드 자원의 연계 장애율은 평균 11.12% 낮은 결과를 얻었다.

Recently, cloud-based resources are stored and used as cloud servers through various devices, but resources transmitted to cloud servers have limitations in flexibility and mobility that connect specific functions to optimize workloads in the cloud. In this paper, we propose a distributed cloud resource allocation technique for minimizing failures in a multi-cloud system that can optimize workloads in the cloud based on factors such as speed, performance, and stability. The proposed technique extracts features of cloud resources by layer and dynamically links and optimizes cloud resource allocation through the convolution process of the patterns of features. In particular, the proposed technique keeps cloud resources consistent by grouping the same pattern of cloud resources into a blockchain to minimize cloud resource allocation in a multi-cloud environment and reducing the size of cloud resources through the convolution process. As a result of the performance evaluation, because the proposed technique is linked to each other according to the probability value between cloud resources, the average cloud resource verification time improved by 7.42%, and the verification delay time decreased by 8.81%. In addition, the average linkage failure rate of cloud resources according to the blockchain group size was 11.12% lower.

인공지능 교육의 방향과 깊이에 대한 논의가 계속되고 있는 오늘날, 본 논문에서는 ｢초·중등 인공지능 교 육 내용기준｣의 고등학교 과정에서 도입되는 CNN(합성곱 신경망)에 대하여 초등학생을 대상으로 교육하는 프로 그램을 제안한다. CNN 알고리즘의 핵심인 합성곱 연산은 곱셈과 덧셈의 연속적인 혼합계산이라는 점에서 수학과 교육과정과의 융합을 모색할 수 있다. 제안된 CNN 교육 프로그램에서 교수자는 합성곱, 스트라이드, 패딩, 풀링이 라는 4개 모듈을 합리적으로 조합하고 학습자의 학년 수준에서 다룰 수 있는 수 체계와 연산을 접목할 수 있다. 권장되는 모듈의 조합은 6가지이며, 초등학교 2학년부터 6학년 학습자까지 학습의 범위와 깊이에 차이를 둔다. 향 후 현장 수업을 통해 교육 프로그램의 효과성을 검증할 필요가 있다.

This paper proposes a program to educate elementary school students about Convolutional Neural Network(CNN). This algorithm introduced in the high school course of the “Elementary and Secondary AI Education Content Standards”. Convolutions, the core of CNN algorithms, are continuous mixed computations of multiplication and addition, so convergence with elementary mathematics curriculum can be sought. In the proposed CNN education program, instructors can combine 4 modules: convolution, stride, padding, and pooling, and incorporate mathematical systems and operations related with the learners' grades. There are 6 recommended combinations of modules, and there are differences in the scope and depth of learning from 2nd to 6th graders. It is necessary to verify the effectiveness of the program through field classes in the future.

우리는 게임 렌더링을 위해 파스텔 드로잉 효과를 시뮬레이션 하는 회선기반 방법을 제안한다. 연구 동기는 굵고 진한 스트로크 패턴, 문지르기 효과, 스트로크 겹침 등 파스텔 드로잉의 속성을 모방하는 것이다. 굵고 진한 파스텔 스트로크 패턴을 위해 우리는 직사각형 모양의 스트로크 패턴 과 유사한 굵은 노이즈를 사용한다. 문지르기 효과는 노이즈를 두 단계로 적분하는 회선 모델을 제안한다: 첫 단계로 회선 방향에 수직한 노이즈를 적분하고, 이를 두 번째 단계에서 회선 방향으 로 적분한다. 스트로크 겹침 효과를 구현하기 위해 우리는 누적 노이즈 필드(ANF, accumulated noise field)를 제안한다. 이는 각 레이어에 노이즈가 분포되어 있다고 가정하고, 서로 다른 레이어 에 있는 스트로크 패턴의 색을 섞는 모델을 의미한다. 이러한 세 가지 개념은 회선기반 프레임 웍 으로 통합되며 시각적으로 만족스러운 게임 장면의 파스텔 드로잉 효과를 보인다.

Many competitive games offer various special rendering effects on their game scenes. Non-photorealistic rendering (NPR) presents fundamental techniques for the special rendering effects. Among various NPR techniques, we focus on a rendering effect that mimics pastel style, which has been rarely studied. We present a convolution-based approach that simulates pastel drawing effects for rendering game scenes. Our motivation is to mimic the properties of pastel drawing artworks such as thick and bold stroke pattern, smearing effect and stroke overlaying. The thick and bold pastel stroke pattern is mimicked by introducing bold noise which resembles a rectangle-shaped stroke pattern. The smearing effect is simulated by devising a convolution model that integrates noise in two-stages: the noise lying in the orthogonal directions of the convolution direction is integrated in the first stage and the results in the first stage are integrated in the second stage. The stroke overlaying is implemented by considering an accumulated noise field (ANF) that considers the distribution of noise in each layer and mimics the blending of colors of the stroke patterns lying in different layer. These three concepts are integrated in a convolution-based framework to present visually pleasing pastel drawing effects for game scene.

본 논문에서는 게임 배경 영상에 대해서 섬세한 일러스트레이션부터 거친 스케치까지 다양한 연필 드로잉 효과를 구현하는 방법을 제안한다. 이 방법은 컨볼루션 알고리즘에 기반한 방법으로 노이즈의 형태를 다양화시켜서 발생시킴으로써 다양한 연필 효과를 추구한다. 기존의 방법에서 발생하는 노이즈는 영상의 톤을 반영하는 픽셀 크기의 입자였다. 본 논문에서는 이러한 픽셀 크기의 노이즈의 형태와 분포를 확장해서 다양한 길이와 두께 및 방향을 갖는 사각형 형태의 노이즈를 사용하는 방법을 고안한다. 그리고 영상 분할 알고리즘을 적용해서 영상의 정보를 추가적으로 활용함으로써 결과 영상의 시각적 완성도를 향상시킨다. 이러한 방법의 장점은 물체, 인물, 장면 등과 같은 다양한 게임 배경 영상에 적용되어 생성된 결과 영상의 우수성을 통해서 입증된다.

In this paper, we present an advanced pencil rendering framework for game wallpaper images that produces a diverse of pencil drawing styles including detailed illustration and rough sketch. Our framework is designed based on a noise generalization scheme that enriches the types of noise which is used in convolution-based pencil rendering framework. The noise used in the conventional framework is a pixel-sized particle whose value reflects the tone of an image. Our scheme extends the shape and the distribution of noise. Instead of using a pixel-sized particle, we use a rectangle-shaped noise whose length, width and orientation are controlled. The performance of the pencil rendering framework is pursued by simplifying the convolution filter. As a result, we produce pencil drawings of various styles. Furthermore, the segmentation algorithm is combined to support local level-of-detail approach that improves the visual attendance of the image. We have proved the excellence of our algorithm by experimenting various game background images including material, portrait, headshot portrait and crime scene. The next step of our framework is to extend the rendering technique to various graphite media.

대장암 검사의 가장 효과적인 진단 방법은 대장내시경 검사이다. 내시경 검사는 소형 카메라를 통하여 확인되는 대 장 내부 영상을 의료인의 육안으로 대장에 돌출한 용종 또는 암으로 성장할 것으로 예측되어지는 용종을 찾아내는 방법이다. 사람의 육안을 통하여 검사가 이루어지는 내시경 검사는 컴퓨터의 영상 학습 기술을 통하여 의료인에게 도움을 제공할 수 있다. 본 연구에서는 내시경 검사를 보조하기 위하여 정상적인 대장, 선종성 용종, 그리고 선암 세 종류로 이루어진 영상 데이터를 분류한다. 제안하는 영상 분류 방법은 심층학습 기반의 영상 분류 기술 중 하나 인 합성곱 신경망(Convolutional Neural Networks) 방법을 통한 내시경 영상 분류 방법을 제안한다. 본 연구에 서 구성한 합성곱 신경망은 총 34개의 합성곱 계층(Convolution Layer)과 하나의 완전연결계층(Fully Connected Layer)을 이룬다. 실험 결과 총 410개의 테스트 데이터에 대해서 94.39%의 인식률을 보였다.

The most effective method of diagnosing colorectal cancer is colonoscopy. colonoscopy is the process of finding protruding or cancerous polyps in a colon via inspection through an internal video of the colon taken with a small camera. The colonoscopy which is usually diagnosed by human visually, can be improved by adapting computer vision learning technology. In this study, classify the images in normal colon, adenomatous polyps, and adenocarcinoma to aid colonoscopy diagnosing. The use of convolutional neural network, one of the image classification techniques based on deep learning, is suggested as the way to classify the colonoscopy images. The convolutional neural network is constructed with 34 convolutional layers and 1 fully-connected layer. The result of this experiment showed a 94.39% accuracy over 410 tests.

To solve the spectrum scarcity problem, dynamic spectrum access (DSA) technology has emerged as a promising solution. Effectively implementing DSA demands accurate and efficient spectrum prediction. However, complex spatiotemporal correlation and heterogeneity in spectrum observations usually make spectral prediction arduous and even ambiguous. In this letter, we propose a spectrum prediction method based on an attention-aided graph convolutional neural network (AttGCN) to capture features in both spatial and temporal dimensions. By leveraging the attention mechanism, the AttGCN adapts its attention weights at different time steps and spatial positions, thus enabling itself to seize changes in spatiotemporal correlations dynamically. Simulation results show that the proposed spectrum prediction method performs better than baseline algorithms in long-term forecasting tasks.

Diabetic Foot Ulcer (DFU) is a complication of diabetes that causes lower limb amputation. In this work, a unique stacked parallel convolution layers-based network (DFU_SPNet) is proposed to perform DFU vs. normal skin classification. The main objective of this work is to design an effective CNN-based classification model, along with proper fine-tuning of optimizer settings. DFU_SPNet consists of 3 blocks of parallel convolution layers with multiple kernel sizes, for local and global feature abstractions. The proposed DFU_SPNet, trained using SGD (with momentum) optimizer with learning rate on the DFUNet dataset, outperformed the current state-of-the-art results with an AUC of 0.974.

In general, it is not a simple task to predict sequences or classify images, and it is even more problematic when both are combined. Nevertheless, biological systems can easily predict sequences and are good at image recognition. For these reasons Long?Short Term Memory and Convolutional Neural Networks were created and were based on the memory and visual systems. These algorithms have shown great properties and shown certain resemblance, yet they are still not the same as their biological counterpart. This article reviews the biological bases and compares them.

Due to the influence of noise in the received signal in non-cooperative communication, it is difficult for existing Automatic modulation classification methods to balance classification accuracy and model complexity. This paper proposes a novel Convolutional Adaptive Noise Reduction (CANR) network, which consists of an Adaptive Noise Reduction (ANR) module and a Convolutional Feature Extraction (CFE) module. The ANR and CFE modules denoise the combined input and capture the spatiotemporal features in the time series. Experiments on benchmark datasets show that the proposed network has the fewest training parameters and state-of-the-art recognition accuracy under the same conditions.

Retinal tissue plays a crucial part in human vision. Infections of retinal tissue and delayed treatment or untreated infection could lead to loss of vision. Additionally, the diagnosis is prone to errors when huge dataset is involved. Therefore, a fully automated model of identification of retinal disease is proposed to reduce human interaction while retaining its high accuracy classification results. This paper introduces an enhanced design of a fully automatic multi-class retina diseases prediction system to assist ophthalmologists in making speedy and accurate investigation. Retinal fundus images, which have been used in this study, were downloaded from the stare website (157 images from five classes: BDR, CRVO, CNV, PDR, and Normal). The five files were categorized according to their annotations conducted by the experienced specialists. The categorized images were first processed with the proposed upgraded contrast-limited adaptive histogram filter for image brightness enhancement, noise reduction, and intensity spectrum normalization. The proposed model was designed with transfer learning method and the fine-tuned pre-trained RESNET50. Eventually, the proposed framework was examined with performance evaluation parameters, recorded a classification rate with 100% sensitivity, 100% specificity, and 100% accuracy. The performance of the proposed model showed a magnificent superiority as compared to the state-of-the-art studies.

Synthetic aperture radar (SAR) is a crucial active imaging technology in remote sensing, offering valuable information for applications like climate monitoring, environmental analysis, and ship surveillance. Ship detection in SAR images remains challenging due to diverse vessel types and environmental interference, especially in inshore areas, despite the proven effectiveness of deep learning-based algorithms. This paper presents an efficient deep learning method named you only look once-shuffle reparameterized blocks with dynamic head (YOLO-SRBD) based on YOLOv8. Additionally, post-processing incorporates the soft non-maximum suppression to enhance precision. Experiments conducted on SAR image datasets demonstrate that the proposed method surpasses the original YOLOv8 both qualitatively and quantitatively, highlighting its feasibility for practical applications. The detection accuracy of the proposed YOLO-SRBD in the high resolution SAR images dataset rose from 89.9% to 91.3%, and the average precision increased from 66.7% to 74.3%, showing significant performance enhancement.

Our objective was to evaluate the CT attenuation coefficient and noise of spatial domain filtering as an alternative to additional image reconstruction using different kernels in abdominal CT. Derived from thin collimated source images was generated using abdomen BIO (very smooth), B20 (smooth), B30 (medium smooth), B40 (medium), B50 (medium sharp), B60 (sharp), B70 (very sharp) and B80 (ultra sharp) kernels. Quantitative CT coefficient and noise measurements provided comparable HU (hounsfield) units in this respect. CT attenuation coefficient (mean HU) values in the abdominal were 60.4 〜 62.2 HU and noise (7.6 〜 63.8 HU) in the liver parenchyma. In the stomach a mean (CT attenuation coefficient) of -2.2 - 0.8 HU and noise (10.1 〜 82.4 HU) was measured. Image reconstructed with a convolution kernel led to an increase in noise, whereas the results for CT attenuation coefficient were comparable. Image medications of image sharpness and noise eliminate the need for reconstruction using different kernels in the future. CT images increase the diagnostic accuracy may be controlled by adjusting CT various kernels, which should be adjusted to take into account the kernels of the CT undeigoing the examination.

주택가격 예측은 개인, 기관, 정부에 이르기까지 다양한 이해관계자들이 관심을 갖는 주요한 주 제이다. 지금까지 통계적 모델을 통해 예측을 해왔지만, 최근 딥러닝 및 머신러닝을 이용하여 예측 하는 연구가 많다. 특히 이미지 정보를 활용한 합성곱신경망(Convolutional Neural Networks)을 적 용한 연구도 진행되고 있다. 그러나 아직 국내에서는 이미지 정보를 이용한 연구는 살펴볼 수 없었 다. 본 연구는 커널밀도추정(Kernel Density Estimation)을 활용하여 인공위성 이미지와 상가 및 편 의시설 밀도 정보를 해당 구역에 표시하고 이를 입력변수로 활용하여 서울시 아파트 가격을 예측하 고자 한다. 본 논문에서는 2021년부터 2022년까지의 주택 거래 기록과 지역소득정보, 인구 밀도, 연령대별 인구 데이터를 수집하였고, 네이버 클라우드 플랫폼의 Static Map API를 통해 인공위성사진을 구역 별로 나누어서 수집했다. 특히 공공데이터포털과 서울 열린데이터 광장에서 수집한 상가, 병원, 공 원, 지하철역, 학교 등 정보는 커널밀도추정을 활용해 밀도 정보로 변환하였다. 이렇게 다양한 형태 로 수집된 데이터는 전처리 과정을 거쳐 면적당 단가로 계산된 아파트 실거래가를 예측하였다. 예 측모델로서는 회귀 모델, 다층 인공신경망 모델, 그리고 합성곱신경망 모델을 이용하였고 이들의 예 측력을 비교하였다. 연구결과는 다음과 같다. 첫째, 회귀 모델 중에서는 인구통계 및 주택관련 변수 에 커널밀도추정으로 구해진 상가 및 편의시설 밀도 정보를 추가한 모델이 예측력이 높았다. 둘째, 다층인공신경망(Multilayer Artificial Neural Network) 모델은 회귀 모델에 비해 상대적으로 높은 성능을 보였다. 셋째, 합성곱신경망과 다층인공신경망을 융합한 모델은 인공위성 이미지 피처, 인구 통계 및 주택 관련 변수, 그리고 커널밀도추정 특성을 모두 사용한 경우 기존 다층 인공신경망과 회귀모델과 비교시 가장 우수한 성능을 보였다. 이 결과는 인공위성 이미지가 주택 가격 예측에 유 용한 정보를 제공하는 것을 시사하며 향후 주택가격예측에 인공위성 이미지, 혹은 커널밀도추정으 로 표시되는 밀도 정보를 합성곱신경망을 이용하는 경우 예측력이 높아졌기에, 향후 이를 개선하는 연구를 기대한다.

House price forecasting is a major topic of interest to a wide range of stakeholders, from individuals to institutions to governments. Until now, predictions have been made using statistical models, but recently, there have been many studies using deep learning and machine learning. In particular, researchers are applying CNN (Convolutional Neural Networks) using image information. However, there have been no studies using image information in Korea. This paper aims to predict apartment prices in Seoul by using Kernel Density Estimation (KDE) to display satellite images and density information of shopping malls and amenities in the area and use them as input variables. In this paper, we collected housing transaction records, local income information, population density, and population data by age group from 2021 to 2022, and collected satellite images divided into zones through Naver Cloud Platform’s Static Map API. In particular, information on shopping malls, hospitals, parks, subway stations, and schools collected from the Open government data portal(www.data.go.kr) and Seoul open data square(data.seoul.go.kr) was converted into density information using KDE. The data collected in various forms was preprocessed to predict the actual transaction price of apartments calculated as a unit price per area. Regression model, multi-layer artificial neural network model, and CNN model were used as prediction models, and their predictive power was compared. The results are as follows. First, among the regression models, the model that adds shopping centre and amenity density information obtained from KDE to demographic and housing-related variables has a high predictive power. Second, the MLP model has a relatively high performance compared to the regression model. Third, the fusion of CNN and MLP performed the best when satellite image features, demographic and housing-related variables, and KDE characteristics were all used, compared to traditional multilayer neural networks and regression models. These results suggest that satellite imagery provides useful information for house price prediction, and we look forward to future research to improve the predictive power of satellite imagery and density information represented by KDE using CNN for house price prediction.

3D object detection is widely applied in robotics and autonomous driving; since 3D scenes in autonomous driving are typically outdoor environments, current methods exhibit substantial computational wastage and significant time delays when using convolution directly in the backbone network. This paper proposes a backbone network based on sparse convolutional spatial-semantic fusion modules to solve this problem. High-level semantic features and low-level spatial features extracted through sub-manifold sparse convolution and sparse convolution are fused to enhance feature representation capabilities. Our proposed backbone network achieves excellent performance on the KITTI dataset.

본 논문에서는 드론을 효과적으로 탐지하기 위해 사용되고 있는 주파수 호핑(Frequency Hopping) 신호 분석 방법과 그 정확도를 비교하였다. 일반적으로 다중 드론의 탐지와 추적을 위해 사용되는 수신 신호의 FFT(Fast Fourier Transform)와 K-평균(K-means) 군집화 기법은 복잡한 운용 환경에서의 간섭 신호와 호핑 시퀀스가 교란(Perturbed)된 상황에서 탐지 정확도가 저하되는 문제가 있다. 본 연구에서는 다수의 드론으로부터 수신되는 신호를 디지털 하향 변환 (Digital Down Conversion), 비교 대상 신호와의 컨볼루션(Convolution) 기반 매칭, 가중치 기반 K-평균(Weighted K-means) 군집화 기법을 사용하여 호핑 신호 탐지 정확도를 비교하였다. 제안된 방법은 낮은 SNR(≤-5dB) 환경에서도 기존 FFT 기법 대비 우수한 탐지 성능을 보였으며, 잡음과 주파수 호핑 시퀀스의 교란(10% perturbed) 상황에서도 우수 한 특성을 나타냈다. 본 연구 결과는 실제 다중 경로 반사 등의 복잡한 드론 운용 환경에서 신뢰성 있는 호핑 신호 탐지 기법으로 활용될 수 있으며, 특히 군용 군집 무인기(UAV) 탐지 및 추적 시스템에 널리 사용될 수 있을 것으로 판단된다.

This paper presents the accuracy of frequency hopping signal analysis techniques used to effectively detect drones. In general, the FFT(Fast Fourier Transform) of the received signals and K-means clustering technique, which are used for multiple drone detection and tracking, have a problem on detection accuracy in complex operating environments with interference signals and perturbed hopping sequences. In this study, we compare the accuracy of hopping signal detection using digital down conversion, convolution-based matching with the reference signals, and weighted K-means clustering techniques for multiple drone signals. The proposed method demonstrates reliable performance in low-SNR(≤-5dB) environments and shows robustness against 10% perturbed hopping signal conditions. The results presented in this study can be utilized as a reliable hopping signal detection technique in such environments involving multipath reflections and other RF interferences. In particular, it is expected to be widely applicable to the military unmanned aerial vehicle (UAV) swarm detection and tracking systems.

Breast cancer remains the foremost cause of cancer-related mortality worldwide. The histopathological diagnosis is impeded by the intricate nature of image interpretation and the presence of inter-observer variability among pathologists. Deep learning (DL) for cancer image understanding has revolutionized accurate breast cancer diagnosis, marking a significant advancement in medical image analysis. Researchers proposed DL-based intelligent models to overcome the challenges of manual observations. However, the existing models suffer from a considerable computational burden, demanding substantial time investments that restrict efficient and scalable breast cancer diagnosis solutions. Our study introduces an automated breast cancer diagnosis system employing a lightweight Convolutional Neural Network (CNN) model, adept at extracting intricate features from histopathological images. Our system has attained superior accuracy through extensive experimentation on a comprehensive breast cancer dataset while employing fewer parameters compared to state-of-theart (SOTA) techniques.

딥러닝을 이용한 객체 검출의 경우 정확도와 실시간성을 모두 요구한다. 그러나, 한정된 자원 환경에서는 수 많은 양의 데이터를 처리하는 딥러닝 모델을 사용하기 어렵다. 이러한 문제 해결을 위해 본 논문에서는 소형 임베디드 장치를 위한 객체 검출을 모델을 제안하였다. 일반적인 검출 모델과 달리 사전 학습된 특징 추출기를 제거한 구조를 사용하여 모델 크기를 최소화하였다. 모델의 구조는 경량화된 컨볼루션 블록을 반복해서 쌓는 구조 로 설계하였다. 또한, 검출 오버헤드를 줄이기 위해 영역 제안 횟수를 크게 줄였다. 제안하는 모델은 공개 데이터 셋인 PASCAL VOC를 사용하여 학습 및 평가하였다. 모델의 정량적 평가를 위해 검출 분야에서 사용하는 average precision으로 검출 성능을 측정하였다. 그리고 실제 임베디드 장치와 유사한 라즈베리 파이에서 검출 속도를 측정하였다. 실험을 통해 기존 검출 방법 대비 향상된 정확도와 빠른 추론 속도를 달성하였다.

In the case of object detection using deep learning, both accuracy and real-time are required. However, it is difficult to use a deep learning model that processes a large amount of data in a limited resource environment. To solve this problem, this paper proposes an object detection model for small embedded devices. Unlike the general detection model, the model size was minimized by using a structure in which the pre-trained feature extractor was removed. The structure of the model was designed by repeatedly stacking lightweight convolution blocks. In addition, the number of region proposals is greatly reduced to reduce detection overhead. The proposed model was trained and evaluated using the public dataset PASCAL VOC. For quantitative evaluation of the model, detection performance was measured with average precision used in the detection field. And the detection speed was measured in a Raspberry Pi similar to an actual embedded device. Through the experiment, we achieved improved accuracy and faster reasoning speed compared to the existing detection method.