Earticle

최근에 생성모델, 특히 ChatGPT와 DALLE-3과 같은 생성모델들은 사회에 큰 영향을 미치며, 이러한 모델들의 잠재공간을 정교하게 조작하고 이해하는 기술이 중요해지고 있다. 본 논문에서는 학 습된 생성모델의 잠재공간을 시각화하여 조작하는 새로운 방법인 “시각화 오토인코더” 방법을 제안한 다. 이 방법은 생성모델의 잠재벡터를 입력으로 하여 3차원 이하의 시각화 가능한 차원으로 축소하여 조작하는 것을 목표로 한다. 본 연구에서는 생성모델의 잠재공간 조작을 “관심 있는 샘플 생성”과 “샘 플 간 변환 과정 시각화” 두 가지로 정의하고, 이를 위해 필요한 수학적 성질 “일대일 대응”과 “locally smoothness”를 가정한 심층신경망 기반의 생성모델로 제한한다. MNIST 데이터셋을 사용하여 실험 한 결과 이러한 성질을 만족하는 심층신경망 모델에 대해 제안된 시각화 오토인코더 방법으로 관심 있는 샘플을 생성, 변환 과정을 시각화할 수 있음을 확인하였다.

Recent generative models, especially generative models such as ChatGPT and DALLE- 3, have had a great impact on society, and technology to intricately manipulate and understand the latent space of these models is becoming important. In this paper, we propose the “Visualization Autoencoder” method, a new method that visualizes and manipulates the latent space of the learned generative model. This method aims to manipulate the latent vectors of the generative model by reducing them to three or less dimensions that can be visualized. In this study, the latent space manipulation of the generative model is defined as two types: “generating samples of interest” and “visualizing the transformation process between samples”. For this purpose, we limit our-selves to a deep neural network-based generative model that assumes the necessary mathematical properties “one-to-one correspondence” and “locally smoothness”. As a result of experiments using the MNIST dataset, we confirmed that the proposed visualization autoencoder method for a deep neural network model that satisfies these properties can generate samples of interest and visualize the conversion process.

심층 신경망(DNN, Deep Neural Network)은 데이터 분석 및 예측에 있어서 높은 효율성으로 인하여 점점 다양한 분야에서 활용되고 있다. 그러나 심층 신경망의 사용이 빈번해짐에 따라, 이와 관련된 보안 위협 요소도 증가하고 있다. 특히, 심층 신경망의 예측에 직접적인 영향을 미칠 수 있는 순전파(forward propagation) 과정과 활성화 함수(activation function)에서 오류가 발생하면, 모델의 예측 정확도에 치명적인 영향을 미칠 수 있다. 본 논문에서는 심층 신경망에서 입력층을 제외한 각 계층으로의 순전파 과정과 출력층에서 사용되는 Softmax 함수에 오류 주입 공격을 수행하고, 그 실험 결과를 분석하였다. MNIST 데이터셋에 대해 글리치 클럭을 이용한 오류 주입 결과, 반복문에 오류를 주입하면 시스템 파라미터에 따른 결정론적 오분류가 일어남을 확인하였다.

The use of Deep Neural Network (DNN) is gradually increasing in various fields due to their high efficiency in data analysis and prediction. However, as the use of deep neural networks becomes more frequent, the security threats associated with them are also increasing. In particular, if a fault occurs in the forward propagation process and activation function that can directly affect the prediction of deep neural network, it can have a fatal damage on the prediction accuracy of the model. In this paper, we performed some fault injection attacks on the forward propagation process of each layer except the input layer in a deep neural network and the Softmax function used in the output layer, and analyzed the experimental results. As a result of fault injection on the MNIST dataset using a glitch clock, we confirmed that faut injection on into the iteration statements can conduct deterministic misclassification depending on the network parameters.

The compaction properties of the soil change depending on the physical properties, and are also affected by crushing of the particles. Since the particle size distribution of soil affects the engineering properties of the soil, it is necessary to analyze the material properties to understand the compaction characteristics. In this study, the size of each sieve was classified into four in the particle size analysis as a material property, and the compaction characteristics were evaluated by multiple regression and maximum dry unit weight. As a result of maximum dry unit weight prediction, multiple regression analysis showed R² of 0.70 or more, and DNN analysis showed R² of 0.80 or more. The reliability of the prediction result analyzed by DNN was evaluated higher than that of multiple regression, and the analysis result of DNN-T showed improved prediction results by 1.87% than DNN. The prediction of maximum dry unit weight using particle size distribution seems to be applied to evaluate the compacting state by identifying the material characteristics of roads and embankments. In addition, the particle size distribution can be used as a parameter for predicting maximum dry unit weight, and it is expected to be of great help in terms of time and cost of applying it to the compaction state evaluation.

Smart transportation, powered by IoT, transforms mobility with interconnected sensors and devices collecting real-time data on traffic, vehicle locations, and passenger needs. This fosters a safer and more sustainable transportation ecosystem, optimizing traffic flow and enhancing public transit efficiency. However, security and privacy challenges emerge in smart transportation systems. Our proposed solution involves a deep neural network (DNN) model trained on extensive datasets from sustainable cities, incorporating historical information like traffic patterns and sensor readings. This model identifies potential malicious nodes, achieving a 90% accuracy rate in predicting threats such as Denial of Service 88%, Whitewash attacks 80%, and Brute Force attacks 75%. This high precision ensures the security and privacy of passenger vehicle data and routes.

Recently, as data demand has increased owing to the rapidly increasing demand for wireless devices and the influence of data traffic, various technologies are being developed to support it. Among them, millimeter-wave (mmWave) frequencies with rich spectra and high data-transmission rates suffer from the problem of large path loss. Accordingly, there is a growing interest in unmanned aerial vehicles (UAVs) and reconfigurable intelligent surfaces (RISs), which can be utilized advantageously to reconstruct wireless communication environments. Therefore, this work considers a large-scale system comprising a number of users and Flying RISs, combining UAVs and RISs to increase algorithm utilization. We propose a deep neural network-based algorithm that places Flying RISs in an appropriate location so that they can support as many users as possible. Simulation results confirmed that the proposed technique could place Flying RISs in an efficient location with higher accuracy and speed in large-scale systems compared to existing techniques.

In this paper, we propose a deep learning approach for solving power allocation problems in massive MIMO networks. We use signal-to-interference-plus-noise-ratio (SINR) and signal-to-leak-plus-noise ratio (SLNR) criteria for linear precoder design to define the max?min and max-prod power allocation challenges. The power allocation process to each user equipment in the base station coverage takes a long time and is inefficient, hence numerous base stations are deployed to serve multiple user equipments. As a result, we develop a deep neural network (DNN) framework in which the user’s equipment position is utilized to train the deep model, which is then used to forecast the ideal power distribution depending on the user’s location. Compared to the traditional optimization approach, the DNN design helps to obtain the optimal solution of the power allocation problem within a short time via a quick-inference process. Simulation results show that the SINR criterion outperforms the SLNR one. Meanwhile, deep learning achieves excellent results in forecasting power allocation with an accuracy of 85% for the max?min strategy and 99% for the max-product approach.

We propose a novel deep neural network model for detecting human activities in untrimmed videos. The process of human activity detection in a video involves two steps: a step to extract features that are effective in recognizing human activities in a long untrimmed video, followed by a step to detect human activities from those extracted features. To extract the rich features from video segments that could express unique patterns for each activity, we employ two different convolutional neural network models, C3D and I-ResNet. For detecting human activities from the sequence of extracted feature vectors, we use BLSTM, a bi-directional recurrent neural network model. By conducting experiments with ActivityNet 200, a large-scale benchmark dataset, we show the high performance of the proposed DeepAct model.

Saliency prediction models are typically trained on natural images, focusing on features such as shape and color. However, predicting saliency in images with text is challenging because the human brain processes text differently than it processes visual objects. To address this research gap, we fine-tuned a saliency model to improve the accuracy of images containing text, specifically, movie posters. Our fine-tuned model?based on GSGNet and TranSalNet?outperformed the original models in predicting the saliency map for movie posters. The experimental results indicate that text elements exhibit patterns that can be learned for better saliency prediction.

With the continuous advancement of computer technology, deep learning models have emerged as innovative tools in shaping various aspects of architectural design. Recognizing the distinctive perspective of children, which differs significantly from that of adults, this paper contends that conventional standards may not always be the most suitable approach in designing urban structures tailored for children. The primary objective of this study is to leverage neural style networks within the design process, specifically adopting the artistic viewpoint found in children's illustrations. By combining the aesthetic paradigm of urban architecture with inspiration drawn from children's aesthetic preferences, the aim is to unearth more creative and subversive aesthetics that challenge traditional norms. The selected context for exploration is the landmark buildings in Qingdao City, Shandong Province, China. Employing the neural style network, the study uses architectural elements of the chosen buildings as content images while preserving their inherent characteristics. The process involves artistic stylization inspired by classic children's illustrations and images from children's picture books. Acting as a conduit for deep learning technology, the research delves into the prospect of seamlessly integrating architectural design styles with the imaginative world of children's illustrations. The outcomes aim to provide fresh perspectives and effective support for the artistic design of contemporary urban buildings.

Capsule endoscopy is one of the increasingly demanded diagnostic methods among patients in recent years because of its ability to observe small intestine difficulties. It is often conducted for 12 to 14 hours, but significant frames constitute only 10% of whole frames. Thus, it has been designed to automatically acquire significant frames through deep learning. For example, studies to track the position of the capsule (stomach, small intestine, etc.) or to extract lesion-related information (polyps, etc.) have been conducted. However, although grouping or labeling the training images according to similar features can improve the performance of a learning model, various attributes (such as degree of wrinkles, presence of valves, etc.) are not considered in conventional approaches. Therefore, we propose a class-labeling method that can be used to design a learning model by constructing a knowledge model focused on main lesions defined in standard terminologies for capsule endoscopy (minimal standard terminology, capsule endoscopy structured terminology). This method enables the designing of a systematic learning model by labeling detailed classes through differentiation of similar characteristics.

In a duck cage, ducks are placed in various states. In particular, if a duck is overturned and falls or dies, it will adversely affect the growing environment. In order to prevent the foregoing, it was necessary to continuously manage the cage for duck growth. This study proposes a method using an object detection algorithm to improve the foregoing. Object detection refers to the work to perform classification and localization of all objects present in the image when an input image is given. To use an object detection algorithm in a duck cage, data to be used for learning should be made and the data should be augmented to secure enough data to learn from. In addition, the time required for object detection and the accuracy of object detection are important. The study collected, processed, and augmented image data for a total of two years in 2021 and 2022 from the duck cage. Based on the objects that must be detected, the data collected as such were divided at a ratio of 9 : 1, and learning and verification were performed. The final results were visually confirmed using images different from the images used for learning. The proposed method is expected to be used for minimizing human resources in the growing process in duck cages and making the duck cages into smart farms.

Conventional split computing approaches for AI models that generate large outputs suffer from long transmission and inference times. Due to the limited resources of the edge server and selfish MDs, some MDs cannot offload their tasks and sacrifice their performance. To address these issues, we formulate an optimization problem to determine one or two split points that minimize inference latency while ensuring fair offloading among MDs. Additionally, we devise a low-complexity heuristic algorithm called fast and fair split computing (F2SC). Evaluation results demonstrate that F2SC reduces inference time by 3.8%~20.1% compared to the conventional approaches while maintaining fairness.

Recognition of banknote serial number is one of the important functions for intelligent banknote counter implementation and can be used for various purposes. However, the previous character recognition method is limited to use due to the font type of the banknote serial number, the variation problem by the solid status, and the recognition speed issue. In this paper, we propose an aspect ratio based character region segmentation and a convolutional neural network (CNN) based banknote serial number recognition method. In order to detect the character region, the character area is determined based on the aspect ratio of each character in the serial number candidate area after the banknote area detection and de-skewing process is performed. Then, we designed and compared four types of CNN models and determined the best model for serial number recognition. Experimental results showed that the recognition accuracy of each character was 99.85%. In addition, it was confirmed that the recognition performance is improved as a result of performing data augmentation. The banknote used in the experiment is Indian rupee, which is badly soiled and the font of characters is unusual, therefore it can be regarded to have good performance. Recognition speed was also enough to run in real time on a device that counts 800 banknotes per minute.

With the development of the sharing economy, existing recommender services are changing from user-item recommendations to user-user recommendations. The most important consideration is that all users should have the best possible satisfaction. To achieve this outcome, the matching service adds information between users and items necessary for the existing recommender service and information between users, so higher-level data mining is required. To this end, this paper proposes a user-to-user matching service (UTU-MS) employing the prediction of mutual satisfaction based on learning. Users were divided into consumers and suppliers, and the properties considered for recommendations were set by filtering and weighting. Based on this process, we implemented a convolutional neural network (CNN)-deep neural network (DNN)-based model that can predict each supplier's satisfaction from the consumer perspective and each consumer's satisfaction from the supplier perspective. After deriving the final mutual satisfaction using the predicted satisfaction, a top recommendation list is recommended to all users. The proposed model was applied to match guests with hosts using Airbnb data, which is a representative sharing economy platform. The proposed model is meaningful in that it has been optimized for the sharing economy and recommendations that reflect user-specific priorities.

For the autoencoder (AE) implemented as a construction component, this paper uses the method of greedy layer-by-layer pre-training without supervision to construct the stacked autoencoder (SAE) to extract the abstract features of the original input data, which is regarded as the input of the logistic regression (LR) model, after which the click-through rate (CTR) of the user to the advertisement under the contextual environment can be obtained. These experiments show that, compared with the usual logistic regression model and support vector regression model used in the field of predicting the advertising CTR in the industry, the SAE-LR model has a relatively large promotion in the AUC value. Based on the improvement of accuracy of advertising CTR prediction, the enterprises can accurately understand and have cognition for the needs of their customers, which promotes the multi-path development with high efficiency and low cost under the condition of internet finance.

In this paper, a vehicle recognition algorithm based on deep convolutional neural network and compression dictionary is proposed. Firstly, the network structure of fine vehicle recognition based on convolutional neural network is introduced. Then, a vehicle recognition system based on multi-scale pyramid convolutional neural network is constructed. The contribution of different networks to the recognition results is adjusted by the adaptive fusion method that adjusts the network according to the recognition accuracy of a single network. The proportion of output in the network output of the entire multiscale network. Then, the compressed dictionary learning and the data dimension reduction are carried out using the effective block structure method combined with very sparse random projection matrix, which solves the computational complexity caused by high-dimensional features and shortens the dictionary learning time. Finally, the sparse representation classification method is used to realize vehicle type recognition. The experimental results show that the detection effect of the proposed algorithm is stable in sunny, cloudy and rainy weather, and it has strong adaptability to typical application scenarios such as occlusion and blurring, with an average recognition rate of more than 95%.

To improve the railway transportation capacity and maximize the benefits of railway transportation, a method for layout optimization of railway transportation route based on deep convolution neural network is proposed in this study. Considering the transportation cost of railway transportation and other factors, the layout model of railway transportation route is constructed. Based on improved ant colony algorithm, the layout model of railway transportation route was optimized, and multiple candidate railway transportation routes were output. Taking into account external information such as regional information, weather conditions and actual information of railway transportation routes, optimization of the candidate railway transportation routes obtained by the improved ant colony algorithm was performed based on deep convolution neural network, and the optimal railway transportation routes were output, and finally layout optimization of railway transportation routes was realized. The experimental results show that the proposed method can obtain the optimal railway transportation route, the shortest transportation length, and the least transportation time, maximizing the interests of railway transportation enterprises.

Trash or garbage is one of the most dangerous health and environmental problems that affect pollution. Pollution affects nature, human life, and wildlife. In this paper, we propose modern solutions for cleaning the environment of trash pollution by enforcing strict action against people who dump trash inappropriately on streets, outside the home, and in unnecessary places. Artificial Intelligence (AI), especially Deep Learning (DL), has been used to automate and solve issues in the world. We availed this as an excellent opportunity to develop a system that identifies trash using a deep convolutional neural network (CNN). This paper proposes a real-time garbage identification system based on a deep CNN architecture with eight distinct classes for the training dataset. After identifying the garbage, the CCTV camera captures a video of the individual placing the trash in the incorrect location and sends an alert notice to the relevant authority.

본 논문에서는 주차장 관리 시스템의 한 부분으로 차량 계수를 위한 감시 카메라 시스템의 컴퓨터 비전과 심층 학습 기반 기법을 제안하고자 한다. You Only Look Once 버전 2 (YOLOv2) 탐지기를 적용하고 YOLOv2 기반의 심층 컨볼루션 신경망(CNN)을 다른 아키텍처와 두 가지 모델로 구성하였다. 제안 된 아키텍처의 효과를 Udacity의 자체 운전 차량 데이터 세트를 사용 하여 설명하였다. 학습 및 테스트 결과, 자동차, 트럭 및 보행자 탐지 시 원래 구조(YOLOv2)의 경우 47.89%의 mAP를 나타내는 것에 비하여, 제안하는 모델의 경우 64.30 %의 mAP를 달성하 여 탐지 정확도가 향상되었음을 증명하였다.

This paper proposes a computer vision and deep learning-based technique for surveillance camera system for vehicle counting as one part of parking lot management system. We applied the You Only Look Once version 2 (YOLOv2) detector and come up with a deep convolutional neural network (CNN) based on YOLOv2 with a different architecture and two models. The effectiveness of the proposed architecture is illustrated using a publicly available Udacity’s self-driving-car datasets. After training and testing, our proposed architecture with new models is able to obtain 64.30% mean average precision which is a better performance compare to the original architecture (YOLOv2) that achieved only 47.89% mean average precision on the detection of car, truck, and pedestrian.