• 한국정보처리학회:학술대회논문집
• /
• 한국정보처리학회 2022년도 추계학술발표대회
• /
• pp.594-596
• /
• 2022

Over the last decades, deep neural networks have demonstrated significant success in various tasks. To address the special vision task, choosing a hot network as backbone to extract feature is a common way in both research and industry project. However, the choice of backbone usually requires the expert experience and affects the performance of the classification task. In this work, we propose a novel idea to support backbone decision-making by exploring the feature attribution and weights distribution of hidden layers from various backbones. We first analyze the visualization of feature maps on different size object and different depth layers to observe learning ability. Then, we compared the variance of weights and feature in last three layers. Based on analysis of the feature and wights, we summarize the traits and commonalities of existing networks.

• 한국군사과학기술학회지
• /
• 제22권2호
• /
• pp.207-214
• /
• 2019

This paper proposes a new feature extraction method for automatically estimating the number of target and detecting the chaff using high range resolution profile(HRRP). Feature of one-dimensional range profile is expected to be limited or missing due to lack of information according to the time. The proposed method considers the dynamic movements of targets depending on the radial velocity. The observed HRRP sequence is used to construct a time-range distribution matrix, then assuming diverse radial velocities reflect the number of target and seduction chaff launch, the proposed method utilizes the characteristic of the gradient distribution on the time-range distribution matrix image, which is validated by electromagnetic computation data and dynamic simulation.

• Genomics & Informatics
• /
• 제18권1호
• /
• pp.7.1-7.7
• /
• 2020

In this paper, we present few technical notes about the distance distribution paradigm for Mosaab-metric using 1, 2, and 3 grams feature extraction techniques to analyze composite data points in high dimensional feature spaces. This technical analysis will help the specialist in bioinformatics and biotechnology to deeply explore the biodiversity of influenza virus genome as a composite data point. Various technical examples are presented in this paper, in addition, the integrated statistical learning pipeline to process segmented genomes of influenza virus is illustrated as sequential-parallel computational pipeline.

• 한국음향학회지
• /
• 제24권4호
• /
• pp.210-216
• /
• 2005

본 논문에서는 잡음에 강인한 음성인식을 위하여 고유공간에 기반을 둔 새로운 특징 정규화 기법을 제안한다. 일반적으로 평균과 분산의 정규화 (MVN)는 켑스트럼 상에서 수행된다. 그러나 최근에 고유공간을 이용한 MVN기법이 소개되었고, 그 고유공간 정규화 기법에서는 하나의 고유공간을 이용하였다. 이 과정에는 켑스트럼 상의 특징 벡터를 선형 주성분 분석 (PCA)행렬을 통하여 고유공간으로 변환시킨 후 MVN을 수행하는 과정이 포함된다. 이 방법에서는 전체 39차의 특징분포를 하나의 고유공간으로 표현하였다. 그러나 이 기법의 경우 전체 특징 분포를 표현함에 세밀함이 떨어지기 때문에 더욱 세밀한 분포의 표현을 위해 본 논문에서는 static 특징, 1차 미분 계수, 2차 미분계수에 각각 유일하고 독립적인 분리된 고유공간을 적용하는 것을 제안하였다. 또한 고유공간에서 정규화 된 훈련 데이터를 이용하여 모델을 만든다. 마지막으로 훈련 데이터의 분포와 잡음환경에서의 테스트 데이터의 분포 특성의 차이를 줄이기 위해 켑스트럼 상에서의 회전 기법을 적용시킨다. 그 결과, 기본적인 고유공간 정규화 기법보다 향상된 성능을 얻을 수 있었다.

• 음성과학
• /
• 제14권2호
• /
• pp.127-135
• /
• 2007

Feature vectors of speech are generally extracted from whole frequency domain. The inherent character of a speaker is located in the low band or high band frequency. However, if the speech is corrupted by narrowband noise with concentrated energy, speaker verification performance is reduced as the individual characteristic is removed. In this paper, we propose a PCA Covariance Model based on the multiband to extract the robust feature vectors against the narrowband noise. First, we divide the overall frequency band into several subbands. Second, the correlation of feature vectors extracted independently from each subband is removed by PCA. The distance obtained from each subband has different distribution. To normalize against the different distribution, we moved the value into the normalized distribution through the mapping function. Finally, the represented value applying the weighting function is used for speaker verification. In the experiments, the proposed method shows better performance of the speaker verification and reduces the computation.

• 복식
• /
• 제62권6호
• /
• pp.200-217
• /
• 2012

The purpose of this research is to examine the specificity of grids to define the characteristics of clothes styles in the Joseon Dynasty period. The significance of examining of the specificity of grids is to find out arbitrary types of the features of grids involved in structuring the Jeogori in the Joseon Dynasty period one by one. The Visual Linguistic Theory was introduced as a methodological tool to exquisitely analyze the characteristics of grids in deep structures of Jeogori in the Joseon Dynasty period. This theory strives to examine sample distribution, the distribution of samples by quality and the distribution of the types of ploidy features. Through the examination, the results are as follows. The grid systems of the Jeogori consisted of diverse proportion systems reaching 86 cases, that is, sequence systems composed of multi-functional, multi-combined bodies. Most ornamental grids had feature angles distributed in a range of $2-20^{\circ}$ that showed a common preference for low sloped diagonal lines or small curvature. Although the preference for certain feature angles were prominent, the feature angles that were used were generally distributed evenly among diverse feature angles to show the characteristics of separation. Therefore, Jeogori makers in the Joseon Dynasty period can be considered as having experimented with many proportion systems to show their aesthetics. In conclusion, based on the results of the examination of feature distributions and related methods to allocate ploidy features, O-type accounted for 66% and thus it was identified that the Jeogori was characterized by O-type. Therefore, it was identified that the characteristic of the Jeogori in the Joseon Dynasty period consisted of O-type fractal structures which are formative structures unique to our nation.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• 제11권8호
• /
• pp.3984-4005
• /
• 2017

Image recognition has become an increasingly important topic for its wide application. It is highly challenging when facing to large-scale database with large variance. The recognition systems rely on a key component, i.e. the low-level feature or the learned mid-level feature. The recognition performance can be potentially improved if the data distribution information is exploited using a more sophisticated way, which usually a function over hidden variable, model parameter and observed data. These methods are called generative score space. In this paper, we propose a discriminative extension for the existing generative score space methods, which exploits class label when deriving score functions for image recognition task. Specifically, we first extend the regular generative models to class conditional models over both observed variable and class label. Then, we derive the mid-level feature mapping from the extended models. At last, the derived feature mapping is embedded into a discriminative classifier for image recognition. The advantages of our proposed approach are two folds. First, the resulted methods take simple and intuitive forms which are weighted versions of existing methods, benefitting from the Bayesian inference of class label. Second, the probabilistic generative modeling allows us to exploit hidden information and is well adapt to data distribution. To validate the effectiveness of the proposed method, we cooperate our discriminative extension with three generative models for image recognition task. The experimental results validate the effectiveness of our proposed approach.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• 제17권12호
• /
• pp.3383-3397
• /
• 2023

Scene graphs serve as semantic abstractions of images and play a crucial role in enhancing visual comprehension and reasoning. However, the performance of Scene Graph Generation is often compromised when working with biased data in real-world situations. While many existing systems focus on a single stage of learning for both feature extraction and classification, some employ Class-Balancing strategies, such as Re-weighting, Data Resampling, and Transfer Learning from head to tail. In this paper, we propose a novel approach that decouples the feature extraction and classification phases of the scene graph generation process. For feature extraction, we leverage a transformer-based architecture and design an adaptive calibration function specifically for predicate classification. This function enables us to dynamically adjust the classification scores for each predicate category. Additionally, we introduce a Distribution Alignment technique that effectively balances the class distribution after the feature extraction phase reaches a stable state, thereby facilitating the retraining of the classification head. Importantly, our Distribution Alignment strategy is model-independent and does not require additional supervision, making it applicable to a wide range of SGG models. Using the scene graph diagnostic toolkit on Visual Genome and several popular models, we achieved significant improvements over the previous state-of-the-art methods with our model. Compared to the TDE model, our model improved mR@100 by 70.5% for PredCls, by 84.0% for SGCls, and by 97.6% for SGDet tasks.

• Journal of Information Processing Systems
• /
• 제14권1호
• /
• pp.176-190
• /
• 2018

This paper presents a novel convolutional neural network based multi-feature fusion learning method for non-rigid 3D model retrieval, which can investigate the useful discriminative information of the heat kernel signature (HKS) descriptor and the wave kernel signature (WKS) descriptor. At first, we compute the 2D shape distributions of the two kinds of descriptors to represent the 3D model and use them as the input to the networks. Then we construct two convolutional neural networks for the HKS distribution and the WKS distribution separately, and use the multi-feature fusion layer to connect them. The fusion layer not only can exploit more discriminative characteristics of the two descriptors, but also can complement the correlated information between the two kinds of descriptors. Furthermore, to further improve the performance of the description ability, the cross-connected layer is built to combine the low-level features with high-level features. Extensive experiments have validated the effectiveness of the designed multi-feature fusion learning method.

• 유통과학연구
• /
• 제19권1호
• /
• pp.75-87
• /
• 2021

Purpose: The objective of this research is to identify the core aspects of persuasive factors influencing consumer's experiential value towards using Augmented Reality (AR) try-on feature while shopping cosmetic products online. The conceptual framework of this study is adopted and integrated from the theoretical study on how narrative experience, media richness, and presence affect the formation of experiential value in the augmented reality interactive technology (ARIT) process. Research design, data and methodology: The sample (n = 550) were collected from online and offline questionnaires by using stratified random sampling and purposive sampling methods. Confirmatory Factor Analysis (CFA) and Structural Equation Model (SEM) were used to analyze the data to confirm goodness-of-fit of the model and hypothesis testing. Results: The results indicated that media richness induced higher experiential value (consumer ROI, playfulness, service excellence and aesthetics), followed by narrative experience and presence towards using AR try-on feature. Conclusions: Consumer's experiential value towards using AR try-on feature when shopping cosmetic products online rely on media richness, narrative experience and presence respectively. Therefore, marketing practitioners are recommended to develop the feature design and content to be more useful, authentic, user-friendly and entertaining to better connect and provide confidence to consumers when shopping cosmetics online.