픽셀 아트는 낮은 해상도와 제한된 색 팔레트를 가지고 영상을 표현한다. 픽셀 아트는 낮은 연 산 성능과 적은 저장 공간을 가지는 초기 컴퓨터 게임에서 주로 사용되었다. 현대에 이르러, 픽셀 아트는 예술이나 퍼즐, 게임과 같은 다양한 분야에서 찾아볼 수 있게 되었다. 본 논문에서는 게임 캐릭터 영상을 입력으로 받는 픽셀 아트 생성 모델을 제안한다. 기존 방법 과는 달리, 합성곱 신경망(CNN:Convolutional-Neural Network)를 픽셀 아트 생성 목적에 맞게 변형하여 이를 이용하는 방법을 제시한다. 기존의 합성곱 연산 후에 upsampling 과정을 추가하여 픽셀 아트가 생성될 수 있도록 하였다. 네트워크는 ground truth와 생성된 픽셀 아트와의 평균 오차 제곱(MSE:Mean Squared Error)을 최소화해나가며 학습을 수행한다. Ground truth는 실제 아티스트가 생성하도록 하였고, 이미지 회전과 반전 기법을 이용하여 augumentation을 수행하였다. 생성된 데이터 집합은 학습, 검증, 시험 데이터로 나누었다. 이러한 데이터 집합을 기반으로 감독 학습을 실시하여 픽셀 아트 생성 네트워크를 학습하였다. 학습 모델의 학습 과정과 학습 정확도를 제시하고, 시험 데이터 뿐만 아니라 다양한 영상에 대한 픽셀 아트 결과도 함께 제시한다.

This paper presents the novel observation model, called Modified Spherical Signature Descriptor(MSSD), capable of representing 2D image generated from 3D point cloud data. The Modified Spherical Signature Descriptor has a uniform mesh grid to accumulate the occupancy evidence caused by neighbor point cloud data. According to a kind of area such as wall, road, tree, car, and so on, the evidence pattern of 2D image looks so different each other. For the parameter learning of Convolutional Neural Network(CNN) layers, these 2D images were applied as the input layer. The Convolutional Neural Network, one of the deep learning methods and familiar with the image analysis, was utilized for the urban structure classification. The case study on CNN practice was introduced in detail in this paper. The simulation results shows that the classification accuracy of CNN with 2D images of the proposed MSSD was improved more than the traditional methods' one.

Deep learning techniques have been studied and developed throughout the medical, agricultural, aviation, and automotive industries. It can be applied to construction fields such as concrete cracks and welding defects. One of the best performing techniques of deep running is CNN technique. CNN means convolutional neural network. In this study, we analyzed crack recognition of sewer with low recognition. Deep learning is generally more accurate with deeper layers, but analysis cost is high. In addition, many variations can occur depending on training options. Therefore, this study performed many parametric studies according to the variations of training options. When analyzed with appropriate training options, the accuracy was over 90% and stable results were obtained

본 연구에서는 아스팔트 콘크리트 도로포장의 표면균열 검출을 위해 합성곱 신경망을 이용하였다. 합성곱 신경망의 학습에 사용되는 표면균열 이미지 데이터의 양에 따른 합성곱 신경망의 성능향상 정도를 평가하였다. 사용된 합성곱 신경망의 구조는 5개의 층으로 구성되 어있으며, 3x3 크기의 convolution filter와 2x2 크기의 pooling kernel을 사용하였다. 합성곱 신경망의 학습을 위해서 도로노면 조사 장비를 통해 구축된 국내 도로포장 표면균열 이미지를 활용하였다. 표면균열 이미지 데이터를 학습한 합성곱 신경망 모델의 표면균열 검출 정확도, 정밀도, 재현율, 미검출율, 과검출율을 평가하였다. 가장 많은 양의 데이터를 학습한 합성곱 신경망 모델의 표면균열 검출 정확도, 정밀도, 재현율은 96.6% 이상, 미검출율, 과검출율은 3.4% 이하의 성능을 나타내었다.

This paper presents a 6-DOF relocalization using a 3D laser scanner and a monocular camera. A relocalization problem in robotics is to estimate pose of sensor when a robot revisits the area. A deep convolutional neural network (CNN) is designed to regress 6-DOF sensor pose and trained using both RGB image and 3D point cloud information in end-to-end manner. We generate the new input that consists of RGB and range information. After training step, the relocalization system results in the pose of the sensor corresponding to each input when a new input is received. However, most of cases, mobile robot navigation system has successive sensor measurements. In order to improve the localization performance, the output of CNN is used for measurements of the particle filter that smooth the trajectory. We evaluate our relocalization method on real world datasets using a mobile robot platform.

3D depth perception has played an important role in robotics, and many sensory methods have also proposed for it. As a photodetector for 3D sensing, single photon avalanche diode (SPAD) is suggested due to sensitivity and accuracy. We have researched for applying a SPAD chip in our fusion system of time-of-fight (ToF) sensor and stereo camera. Our goal is to upsample of SPAD resolution using RGB stereo camera. Currently, we have 64 x 32 resolution SPAD ToF Sensor, even though there are higher resolution depth sensors such as Kinect V2 and Cube-Eye. This may be a weak point of our system, however we exploit this gap using a transition of idea. A convolution neural network (CNN) is designed to upsample our low resolution depth map using the data of the higher resolution depth as label data. Then, the upsampled depth data using CNN and stereo camera depth data are fused using semi-global matching (SGM) algorithm. We proposed simplified fusion method created for the embedded system.

최근의 인공 신경망(Neural Network) 기법은 전통적인 분류 문제와 군집화 문제 해결에서 벗어나 이미지 생성 같은 컨텐츠 생성에서도 좋은 성능을 보이고 있다. 본 연구에서는 차세대 컨텐츠 생성 기법으로 인공신경망을 이용한 이미지 생성기법을 제안한다. 제안하는 인공신경망 모델은 두 개의 이미지를 입력받아서 하나의 이미지에서는 색상을, 다른 이미지에서는 모양을 가져와 새로운 이미지로 조합해낸다. 이 모델은 컨볼루션 인공신경망(Convolutional Neural Network)으로 제작되 었으며 각각 이미지에서 색상과 모양을 추출해내는 두 개의 인코더와 각 인코더의 값을 모두 넘겨 받아 하나의 조합이 되는 이미지를 생성해내는 하나의 디코더로 구성이 되어있다. 본 연구의 성과는 저비용으로 게임 개발 프로세스 상 다양한 2차원 이미지 생성 및 보정 작업에 활용될 수 있다.

This paper presents a convolutional neural network to automatically conduct the peak picking in frequency domain of structural responses. The peaks in frequency domain have a high potential to be the natural frequencies, which are one of the important indicator to be used for structural health monitoring purposes, such as damage detection, cable tension estimation, and finite element model updating. In general, the peaks with the corresponding natural frequencies are manually selected by the users from the frequency domain. Although this previous approach is possible to simply extract the candidate of natural frequencies, it is inappropriate in the practical applications of the long-term monitoring and the implementation for wireless smart sensor. To overcome the drawbacks, this study proposes the convolutional neural network that can automatically identify the peaks with the corresponding natural frequencies from the frequency domain of structural responses.

This paper proposes a convolutional neural network model for distinguishing areas occupied by obstacles from a LiDAR image converted from a 3D point cloud. The channels of a LiDAR image used as input consist of the distances to 3D points, the reflectivities of 3D points, and the heights of 3D points from the ground. The proposed model uses a LiDAR image as an input and outputs a result of a segmented LiDAR image. The proposed model adopts refinement modules with skip connections to segment a LiDAR image. The refinement modules with skip connections in the proposed model make it possible to construct a complex structure with a small number of parameters than a convolutional neural network model with a linear structure. Using the proposed model, it is possible to distinguish areas in a LiDAR image occupied by obstacles such as vehicles, pedestrians, and bicyclists. The proposed model can be applied to recognize surrounding obstacles and to search for safe paths.

In this paper, we propose a jellyfish distribution recognition and monitoring system using a UAV (unmanned aerial vehicle). The UAV was designed to satisfy the requirements for flight in ocean environment. The target jellyfish, Aurelia aurita, is recognized through convolutional neural network and its distribution is calculated. The modified deep neural network architecture has been developed to have reliable recognition accuracy and fast operation speed. Recognition speed is about 400 times faster than GoogLeNet by using a lightweight network architecture. We also introduce the method for selecting candidates to be used as inputs to the proposed network. The recognition accuracy of the jellyfish is improved by removing the probability value of the meaningless class among the probability vectors of the evaluated input image and re-evaluating it by normalization. The jellyfish distribution is calculated based on the unit jellyfish image recognized. The distribution level is defined by using the novelty concept of the distribution map buffer.

As drones gain more popularity these days, drone detection becomes more important part of the drone systems for safety, privacy, crime prevention and etc. However, existing drone detection systems are expensive and heavy so that they are only suitable for industrial or military purpose. This paper proposes a novel approach for training Convolutional Neural Networks to detect drones from images that can be used in embedded systems. Unlike previous works that consider the class probability of the image areas where the class object exists, the proposed approach takes account of all areas in the image for robust classification and object detection. Moreover, a novel loss function is proposed for the CNN to learn more effectively from limited amount of training data. The experimental results with various drone images show that the proposed approach performs efficiently in real drone detection scenarios.