• Journal of the Korea Institute of Military Science and Technology
• /
• v.13 no.4
• /
• pp.683-688
• /
• 2010

The ability to navigate autonomously in off-road terrain is the most critical technology needed for Unmanned Ground Vehicles(UGV). In this paper, we present a method for vision-based terrain cover classification using DCT features. To classify the terrain, we acquire image from a CCD sensor, then the image is divided into fixed size of blocks. And each block transformed into DCT image then extracts features which reflect frequency band characteristics. Neural network classifier is used to classify the features. The proposed method is validated and verified through many experiments and we compare it with wavelet feature based method. The results show that the proposed method is more efficiently classify the terrain-cover than wavelet feature based one.

• Proceedings of the Korean Information Science Society Conference
• /
• 2004.04b
• /
• pp.796-798
• /
• 2004

기존의 번호판 인식 시스템에서는 대부분 카메라가 고정 상태에서 차량의 전면부를 찍어 영상을 획득하고, 이로부터 번호판을 추출하고 인식한다 그러나 본 연구에서는 기존 연구들과 달리 이동 중인 자동차에 카메라를 설치하여 움직이는 자동차의 영상을 획득하여 번호판을 추출하고 인식한다. 인식하고자 하는 영상이 잡음이나 왜곡 없이 깨끗하다면 인식 과정은 간단하게 수행될 것이다. 그러나, 실제로 얻어진 영상은 간단한 방법으로 인식하기에는 어려올 정도로 왜곡이나 변형이 심한 경우가 많다. 따라서 본 논문에서는 SVM 전단에 모듈라 신경망을 결합하여 인식하는 방법을 사용함으로써 잡음과 같은 변형에 덜 민감하도록 하고자 하였다. 실험결과, 제안하는 분류기를 이용한 방법이 번호판 인식에 우수한 성능을 보임을 확인하였다.

• Journal of Korea Multimedia Society
• /
• v.25 no.6
• /
• pp.769-782
• /
• 2022

A study on agricultural automation became more important. In Korea, sweet persimmon farmers spend a lot of time and effort on classifying profitable persimmons. In this paper, we propose and implement an efficient grading algorithm for persimmons before shipment. We gathered more than 1,750 images of persimmons, and the images were graded and labeled for classifications purpose. Our main algorithm is based on EfficientDet object detection model but we implemented more exquisite method for better classification performance. In order to improve the precision of classification, we adopted a machine learning algorithm, which was proposed by PyCaret machine learning workflow generation library. Finally we acquired an improved classification model with the accuracy score of 81%.

• ETRI Journal
• /
• v.44 no.2
• /
• pp.327-338
• /
• 2022

Classification models for human action recognition require robust features and large training sets for good generalization. However, data augmentation methods are employed for imbalanced training sets to achieve higher accuracy. These samples generated using data augmentation only reflect existing samples within the training set, their feature representations are less diverse and hence, contribute to less precise classification. This paper presents new data augmentation and action representation approaches to grow training sets. The proposed approach is based on two fundamental concepts: virtual video generation for augmentation and representation of the action videos through robust features. Virtual videos are generated from the motion history templates of action videos, which are convolved using a convolutional neural network, to generate deep features. Furthermore, by observing an objective function of the genetic algorithm, the spatiotemporal features of different samples are combined, to generate the representations of the virtual videos and then classified through an extreme learning machine classifier on MuHAVi-Uncut, iXMAS, and IAVID-1 datasets.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.16 no.11
• /
• pp.3658-3679
• /
• 2022

Classification of persons wearing and not wearing face masks in images has emerged as a new computer vision problem during the COVID-19 pandemic. In order to address this problem and scale up the research in this domain, in this paper a hybrid technique by employing ResNet-101 and multi-layer perceptron (MLP) classifier has been proposed. The proposed technique is tested and validated on a self-created face masks classification dataset and a standard dataset. On self-created dataset, the proposed technique achieved a classification accuracy of 97.3%. To embrace the proposed technique, six other state-of-the-art CNN feature extractors with six other classical machine learning classifiers have been tested and compared with the proposed technique. The proposed technique achieved better classification accuracy and 1-6% higher precision, recall, and F1 score as compared to other tested deep feature extractors and machine learning classifiers.

• Korean Journal of Geomatics
• /
• v.3 no.1
• /
• pp.33-41
• /
• 2003

Every year, several typhoons hit the Korean peninsula and cause severe damage. For the prevention and accurate estimation of these damages, real time or almost real time flood information is essential. Because of weather conditions, images taken by optic sensors or LIDAR are sometimes not appropriate for an accurate estimation of water areas during typhoon. In this case SAR (Synthetic Aperture Radar) images which are independent of weather condition can be useful for the estimation of flood areas. To get detailed information about floods from satellite imagery, accurate classification of water areas is the most important step. A commonly- and widely-used classification methods is the ML(Maximum Likelihood) method which assumes that the distribution of brightness values of the images follows a Gaussian distribution. The distribution of brightness values of the SAR image, however, usually does not follow a Gaussian distribution. For this reason, in this study the ANN (Artificial Neural Networks) method independent of the statistical characteristics of images is applied to the SAR imagery. RADARS A TSAR images are primarily used for extraction of water areas, and DEM (Digital Elevation Model) is used as supplementary data to evaluate the ground undulation effect. Water areas are also extracted from KOMPSAT image achieved by optic sensors for comparison purpose. Both ANN and ML methods are applied to flat and mountainous areas to extract water areas. The estimated areas from satellite imagery are compared with those of manually extracted results. As a result, the ANN classifier performs better than the ML method when only the SAR image was used as input data, except for mountainous areas. When DEM was used as supplementary data for classification of SAR images, there was a 5.64% accuracy improvement for mountainous area, and a similar result of 0.24% accuracy improvement for flat areas using artificial neural networks.

• Journal of the Institute of Electronics Engineers of Korea CI
• /
• v.39 no.3
• /
• pp.44-52
• /
• 2002

In this paper, we present an analytical method for decision boundary feature extraction for neural networks. It has been shown that all the features necessary to achieve the same classification accuracy xxxas in the original space can be obtained from the vectors normal to decision boundaries. However, the vector normal to the decision boundary of a neural network has been calculated numerically using a gradient approximation. This process is time-consuming and the normal vector may be inaccurately estimated. In this paper, we propose a method to improve the performance of the previous decision boundary feature extraction for neural networks by accurately calculating the normal vector When the normal vectors are computed analytically, it is possible to reduce the processing time significantly and improve the performance of the previous implementation that employs numerical approximation.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.13 no.5
• /
• pp.2674-2697
• /
• 2019

Identity authentication is a crucial line of defense for network security, and passwords are still the mainstream of identity authentication. So far trawling password attacking has been extensively studied, but the research related with personal information is always sporadic. Probabilistic context-free grammar (PCFG) and Markov chain-based models perform greatly well in trawling guessing. In this paper we propose a systematic targeted attacking model based on structure partition and string reorganization by migrating the above two models to targeted attacking, denoted as TG-SPSR. In structure partition phase, besides dividing passwords to basic structure similar to PCFG, we additionally define a trajectory-based keyboard pattern in the basic grammar and introduce index bits to accurately characterize the position of special characters. Moreover, we also construct a BiLSTM recurrent neural network classifier to characterize the behavior of password reuse and modification after defining nine kinds of modification rules. Extensive experimental results indicate that in online attacking, TG-SPSR outperforms traditional trawling attacking algorithms by average about 275%, and respectively outperforms its foremost counterparts, Personal-PCFG, TarGuess-I, by about 70% and 19%; In offline attacking, TG-SPSR outperforms traditional trawling attacking algorithms by average about 90%, outperforms Personal-PCFG and TarGuess-I by 85% and 30%, respectively.

• Journal of Biomedical Engineering Research
• /
• v.19 no.1
• /
• pp.1-8
• /
• 1998

Clustered microcalcifications on X-ray mammograms are an important sign for early detection of breast cancer. This paper proposes a computer-aided diagnosis method for the detection of clustered microcalcifications and marking their locations on digitized mammograms. The proposed detection method consists of the region of interest (ROI) selection, the film-artifact removal, the surrounding texture analysis method for the detection of clustered microcalcifications, which is based on the second-order histogram in two nested surrounding regions on the current pixel. This paper also describes the effectiveness of the proposed film-artifact removal filter in terms of the classification performance with the receiver operating-characteristics(ROC) analysis. A three-layer backpropagation neural network is employed as a classifier. The appropriate marking for the locations of clustered microcalcifications can be used to alert radiologists to locations of suspicious lesions.

• The Journal of Korean Institute of Electromagnetic Engineering and Science
• /
• v.22 no.10
• /
• pp.981-989
• /
• 2011

In this paper, we performed radar target classification for the five scale models using ultra-wideband measured signal. In order to compare the performance, the 2 GHz(2~4 GHz), 4 GHz(2~6 GHz), and 6 GHz(2~8 GHz) bandwidth were used. Short time Fourier transform(STFT) and continuous wavelet transform(CWT) are used for target feature extraction. Extracted feature vectors are used as input for the multi-layerd perceptron(MLP) neural network classifier. The results show that as the bandwidth is wider, the performance is better.