• Journal of Intelligence and Information Systems
• /
• v.24 no.3
• /
• pp.1-19
• /
• 2018

Large amount of data is now available for research and business sectors to extract knowledge from it. This data can be in the form of unstructured data such as audio, text, and image data and can be analyzed by deep learning methodology. Deep learning is now widely used for various estimation, classification, and prediction problems. Especially, fashion business adopts deep learning techniques for apparel recognition, apparel search and retrieval engine, and automatic product recommendation. The core model of these applications is the image classification using Convolutional Neural Networks (CNN). CNN is made up of neurons which learn parameters such as weights while inputs come through and reach outputs. CNN has layer structure which is best suited for image classification as it is comprised of convolutional layer for generating feature maps, pooling layer for reducing the dimensionality of feature maps, and fully-connected layer for classifying the extracted features. However, most of the classification models have been trained using online product image, which is taken under controlled situation such as apparel image itself or professional model wearing apparel. This image may not be an effective way to train the classification model considering the situation when one might want to classify street fashion image or walking image, which is taken in uncontrolled situation and involves people's movement and unexpected pose. Therefore, we propose to train the model with runway apparel image dataset which captures mobility. This will allow the classification model to be trained with far more variable data and enhance the adaptation with diverse query image. To achieve both convergence and generalization of the model, we apply Transfer Learning on our training network. As Transfer Learning in CNN is composed of pre-training and fine-tuning stages, we divide the training step into two. First, we pre-train our architecture with large-scale dataset, ImageNet dataset, which consists of 1.2 million images with 1000 categories including animals, plants, activities, materials, instrumentations, scenes, and foods. We use GoogLeNet for our main architecture as it has achieved great accuracy with efficiency in ImageNet Large Scale Visual Recognition Challenge (ILSVRC). Second, we fine-tune the network with our own runway image dataset. For the runway image dataset, we could not find any previously and publicly made dataset, so we collect the dataset from Google Image Search attaining 2426 images of 32 major fashion brands including Anna Molinari, Balenciaga, Balmain, Brioni, Burberry, Celine, Chanel, Chloe, Christian Dior, Cividini, Dolce and Gabbana, Emilio Pucci, Ermenegildo, Fendi, Giuliana Teso, Gucci, Issey Miyake, Kenzo, Leonard, Louis Vuitton, Marc Jacobs, Marni, Max Mara, Missoni, Moschino, Ralph Lauren, Roberto Cavalli, Sonia Rykiel, Stella McCartney, Valentino, Versace, and Yve Saint Laurent. We perform 10-folded experiments to consider the random generation of training data, and our proposed model has achieved accuracy of 67.2% on final test. Our research suggests several advantages over previous related studies as to our best knowledge, there haven't been any previous studies which trained the network for apparel image classification based on runway image dataset. We suggest the idea of training model with image capturing all the possible postures, which is denoted as mobility, by using our own runway apparel image dataset. Moreover, by applying Transfer Learning and using checkpoint and parameters provided by Tensorflow Slim, we could save time spent on training the classification model as taking 6 minutes per experiment to train the classifier. This model can be used in many business applications where the query image can be runway image, product image, or street fashion image. To be specific, runway query image can be used for mobile application service during fashion week to facilitate brand search, street style query image can be classified during fashion editorial task to classify and label the brand or style, and website query image can be processed by e-commerce multi-complex service providing item information or recommending similar item.

• Journal of Intelligence and Information Systems
• /
• v.25 no.1
• /
• pp.163-177
• /
• 2019

As smartphones are getting widely used, human activity recognition (HAR) tasks for recognizing personal activities of smartphone users with multimodal data have been actively studied recently. The research area is expanding from the recognition of the simple body movement of an individual user to the recognition of low-level behavior and high-level behavior. However, HAR tasks for recognizing interaction behavior with other people, such as whether the user is accompanying or communicating with someone else, have gotten less attention so far. And previous research for recognizing interaction behavior has usually depended on audio, Bluetooth, and Wi-Fi sensors, which are vulnerable to privacy issues and require much time to collect enough data. Whereas physical sensors including accelerometer, magnetic field and gyroscope sensors are less vulnerable to privacy issues and can collect a large amount of data within a short time. In this paper, a method for detecting accompanying status based on deep learning model by only using multimodal physical sensor data, such as an accelerometer, magnetic field and gyroscope, was proposed. The accompanying status was defined as a redefinition of a part of the user interaction behavior, including whether the user is accompanying with an acquaintance at a close distance and the user is actively communicating with the acquaintance. A framework based on convolutional neural networks (CNN) and long short-term memory (LSTM) recurrent networks for classifying accompanying and conversation was proposed. First, a data preprocessing method which consists of time synchronization of multimodal data from different physical sensors, data normalization and sequence data generation was introduced. We applied the nearest interpolation to synchronize the time of collected data from different sensors. Normalization was performed for each x, y, z axis value of the sensor data, and the sequence data was generated according to the sliding window method. Then, the sequence data became the input for CNN, where feature maps representing local dependencies of the original sequence are extracted. The CNN consisted of 3 convolutional layers and did not have a pooling layer to maintain the temporal information of the sequence data. Next, LSTM recurrent networks received the feature maps, learned long-term dependencies from them and extracted features. The LSTM recurrent networks consisted of two layers, each with 128 cells. Finally, the extracted features were used for classification by softmax classifier. The loss function of the model was cross entropy function and the weights of the model were randomly initialized on a normal distribution with an average of 0 and a standard deviation of 0.1. The model was trained using adaptive moment estimation (ADAM) optimization algorithm and the mini batch size was set to 128. We applied dropout to input values of the LSTM recurrent networks to prevent overfitting. The initial learning rate was set to 0.001, and it decreased exponentially by 0.99 at the end of each epoch training. An Android smartphone application was developed and released to collect data. We collected smartphone data for a total of 18 subjects. Using the data, the model classified accompanying and conversation by 98.74% and 98.83% accuracy each. Both the F1 score and accuracy of the model were higher than the F1 score and accuracy of the majority vote classifier, support vector machine, and deep recurrent neural network. In the future research, we will focus on more rigorous multimodal sensor data synchronization methods that minimize the time stamp differences. In addition, we will further study transfer learning method that enables transfer of trained models tailored to the training data to the evaluation data that follows a different distribution. It is expected that a model capable of exhibiting robust recognition performance against changes in data that is not considered in the model learning stage will be obtained.

• Journal of Intelligence and Information Systems
• /
• v.27 no.1
• /
• pp.83-102
• /
• 2021

The government recently announced various policies for developing big-data and artificial intelligence fields to provide a great opportunity to the public with respect to disclosure of high-quality data within public institutions. KSURE(Korea Trade Insurance Corporation) is a major public institution for financial policy in Korea, and thus the company is strongly committed to backing export companies with various systems. Nevertheless, there are still fewer cases of realized business model based on big-data analyses. In this situation, this paper aims to develop a new business model which can be applied to an ex-ante prediction for the likelihood of the insurance accident of credit guarantee. We utilize internal data from KSURE which supports export companies in Korea and apply machine learning models. Then, we conduct performance comparison among the predictive models including Logistic Regression, Random Forest, XGBoost, LightGBM, and DNN(Deep Neural Network). For decades, many researchers have tried to find better models which can help to predict bankruptcy since the ex-ante prediction is crucial for corporate managers, investors, creditors, and other stakeholders. The development of the prediction for financial distress or bankruptcy was originated from Smith(1930), Fitzpatrick(1932), or Merwin(1942). One of the most famous models is the Altman's Z-score model(Altman, 1968) which was based on the multiple discriminant analysis. This model is widely used in both research and practice by this time. The author suggests the score model that utilizes five key financial ratios to predict the probability of bankruptcy in the next two years. Ohlson(1980) introduces logit model to complement some limitations of previous models. Furthermore, Elmer and Borowski(1988) develop and examine a rule-based, automated system which conducts the financial analysis of savings and loans. Since the 1980s, researchers in Korea have started to examine analyses on the prediction of financial distress or bankruptcy. Kim(1987) analyzes financial ratios and develops the prediction model. Also, Han et al.(1995, 1996, 1997, 2003, 2005, 2006) construct the prediction model using various techniques including artificial neural network. Yang(1996) introduces multiple discriminant analysis and logit model. Besides, Kim and Kim(2001) utilize artificial neural network techniques for ex-ante prediction of insolvent enterprises. After that, many scholars have been trying to predict financial distress or bankruptcy more precisely based on diverse models such as Random Forest or SVM. One major distinction of our research from the previous research is that we focus on examining the predicted probability of default for each sample case, not only on investigating the classification accuracy of each model for the entire sample. Most predictive models in this paper show that the level of the accuracy of classification is about 70% based on the entire sample. To be specific, LightGBM model shows the highest accuracy of 71.1% and Logit model indicates the lowest accuracy of 69%. However, we confirm that there are open to multiple interpretations. In the context of the business, we have to put more emphasis on efforts to minimize type 2 error which causes more harmful operating losses for the guaranty company. Thus, we also compare the classification accuracy by splitting predicted probability of the default into ten equal intervals. When we examine the classification accuracy for each interval, Logit model has the highest accuracy of 100% for 0~10% of the predicted probability of the default, however, Logit model has a relatively lower accuracy of 61.5% for 90~100% of the predicted probability of the default. On the other hand, Random Forest, XGBoost, LightGBM, and DNN indicate more desirable results since they indicate a higher level of accuracy for both 0~10% and 90~100% of the predicted probability of the default but have a lower level of accuracy around 50% of the predicted probability of the default. When it comes to the distribution of samples for each predicted probability of the default, both LightGBM and XGBoost models have a relatively large number of samples for both 0~10% and 90~100% of the predicted probability of the default. Although Random Forest model has an advantage with regard to the perspective of classification accuracy with small number of cases, LightGBM or XGBoost could become a more desirable model since they classify large number of cases into the two extreme intervals of the predicted probability of the default, even allowing for their relatively low classification accuracy. Considering the importance of type 2 error and total prediction accuracy, XGBoost and DNN show superior performance. Next, Random Forest and LightGBM show good results, but logistic regression shows the worst performance. However, each predictive model has a comparative advantage in terms of various evaluation standards. For instance, Random Forest model shows almost 100% accuracy for samples which are expected to have a high level of the probability of default. Collectively, we can construct more comprehensive ensemble models which contain multiple classification machine learning models and conduct majority voting for maximizing its overall performance.

• Journal of the Korean Institute of Landscape Architecture
• /
• v.51 no.3
• /
• pp.70-82
• /
• 2023

The recent surge of IT and data acquisition is shifting the paradigm in all aspects of life, and these advances are also affecting academic fields. Research topics and methods are being improved through academic exchange and connections. In particular, data-based research methods are employed in various academic fields, including landscape architecture, where continuous research is needed. Therefore, this study aims to investigate the possibility of developing a landscape preference evaluation and prediction model using machine learning, a branch of Artificial Intelligence, reflecting the current situation. To achieve the goal of this study, machine learning techniques were applied to the landscaping field to build a landscape preference evaluation and prediction model to verify the simulation accuracy of the model. For this, wind power facility landscape images, recently attracting attention as a renewable energy source, were selected as the research objects. For analysis, images of the wind power facility landscapes were collected using web crawling techniques, and an analysis dataset was built. Orange version 3.33, a program from the University of Ljubljana was used for machine learning analysis to derive a prediction model with excellent performance. IA model that integrates the evaluation criteria of machine learning and a separate model structure for the evaluation criteria were used to generate a model using kNN, SVM, Random Forest, Logistic Regression, and Neural Network algorithms suitable for machine learning classification models. The performance evaluation of the generated models was conducted to derive the most suitable prediction model. The prediction model derived in this study separately evaluates three evaluation criteria, including classification by type of landscape, classification by distance between landscape and target, and classification by preference, and then synthesizes and predicts results. As a result of the study, a prediction model with a high accuracy of 0.986 for the evaluation criterion according to the type of landscape, 0.973 for the evaluation criterion according to the distance, and 0.952 for the evaluation criterion according to the preference was developed, and it can be seen that the verification process through the evaluation of data prediction results exceeds the required performance value of the model. As an experimental attempt to investigate the possibility of developing a prediction model using machine learning in landscape-related research, this study was able to confirm the possibility of creating a high-performance prediction model by building a data set through the collection and refinement of image data and subsequently utilizing it in landscape-related research fields. Based on the results, implications, and limitations of this study, it is believed that it is possible to develop various types of landscape prediction models, including wind power facility natural, and cultural landscapes. Machine learning techniques can be more useful and valuable in the field of landscape architecture by exploring and applying research methods appropriate to the topic, reducing the time of data classification through the study of a model that classifies images according to landscape types or analyzing the importance of landscape planning factors through the analysis of landscape prediction factors using machine learning.

• Proceedings of the Korean Society of Near Infrared Spectroscopy Conference
• /
• 2001.06a
• /
• pp.1032-1032
• /
• 2001

Artificial Neural Network (ANN) calibration techniques have been used commercially for agricultural applications since the mid-nineties. Global models, based on transmission data from 850 to 1050 nm, are used routinely to measure protein and moisture in wheat and barley and also moisture in triticale, rye, and oats. These models are currently used commercially in approx. 15 countries throughout the world. Results concerning earlier European ANN models are being published elsewhere. Some of the findings from that study will be discussed here. ANN models have also been developed for coarsely ground samples of compound feed and feed ingredients, again measured in transmission mode from 850 to 1050 nm. The performance of models for pig- and poultry feed will be discussed briefly. These models were developed from a very large data set (more than 20,000 records), and cover a very broad range of finished products. The prediction curves are linear over the entire range for protein, fat moisture, fibre, and starch (measured only on poultry feed), and accuracy is in line with the performance of smaller models based on Partial Least Squares (PLS). A simple bias adjustment is sufficient for calibration transfer across instruments. Recently, we have investigated the possible use of ANN for a different type of NIR spectrometer, based on reflectance data from 1100 to 2500 nm. In one study, based on data for protein, fat, and moisture measured on unground compound feed samples, dedicated ANN models for specific product classes (cattle feed, pig feed, broiler feed, and layers feed) gave moderately better Standard Errors of Prediction (SEP) compared to modified PLS (MPLS). However, if the four product classes were combined into one general calibration model, the performance of the ANN model deteriorated only slightly compared to the class-specific models, while the SEP values for the MPLS predictions doubled. Brix value in molasses is a measure of sugar content. Even with a huge dataset, PLS models were not sufficiently accurate for commercial use. In contrast an ANN model based on the same data improved the accuracy considerably and straightened out non-linearity in the prediction plot. The work of Mr. David Funk (GIPSA, U. S. Department of Agriculture) who has studied the influence of various types of spectral distortions on ANN- and PLS models, thereby providing comparative information on the robustness of these models towards instrument differences, will be discussed. This study was based on data from different classes of North American wheat measured in transmission from 850 to 1050 nm. The distortions studied included the effect of absorbance offset pathlength variation, presence of stray light bandwidth, and wavelength stretch and offset (either individually or combined). It was shown that a global ANN model was much less sensitive to most perturbations than class-specific GIPSA PLS calibrations. It is concluded that ANN models based on large data sets offer substantial advantages over PLS models with respect to accuracy, range of materials that can be handled by a single calibration, stability, transferability, and sensitivity to perturbations.

• Journal of KIISE:Software and Applications
• /
• v.35 no.9
• /
• pp.572-587
• /
• 2008

The Self-organizing Feature Map(SOFM) that is one of unsupervised neural networks is a very powerful tool for data clustering and visualization in high-dimensional data sets. Although the SOFM has been applied in many engineering problems, it needs to cluster similar weights into one class on the trained SOFM as a post-processing, which is manually performed in many cases. The traditional clustering algorithms, such as t-means, on the trained SOFM however do not yield satisfactory results, especially when clusters have arbitrary shapes. This paper proposes automatic clustering on trained SOFM, which can deal with arbitrary cluster shapes and be globally optimized by graph cuts. When using the graph cuts, the graph must have two additional vertices, called terminals, and weights between the terminals and vertices of the graph are generally set based on data manually obtained by users. The Proposed method automatically sets the weights based on mode-seeking on a distance matrix. Experimental results demonstrated the effectiveness of the proposed method in texture segmentation. In the experimental results, the proposed method improved precision rates compared with previous traditional clustering algorithm, as the method can deal with arbitrary cluster shapes based on the graph-theoretic clustering.

• Journal of the Korea Academia-Industrial cooperation Society
• /
• v.21 no.7
• /
• pp.285-291
• /
• 2020

In manufacturing, humans are being replaced with robots, but expert skills remain difficult to convert to data, making them difficult to apply to industrial robots. One method is by visual motion recognition, but physical features may be judged differently depending on the image data. This study aimed to improve the accuracy of vision methods for estimating the posture of humans. Three OpenPose vision models were applied: MPII, COCO, and COCO+foot. To identify the effects of face-covering accessories and image preprocessing on the Convolutional Neural Network (CNN) structure, the presence/non-presence of accessories, image size, and filtering were set as the parameters affecting the identification of a human's posture. For each parameter, image data were applied to the three models, and the errors between the actual and predicted values, as well as the percentage correct keypoints (PCK), were calculated. The COCO+foot model showed the lowest sensitivity to all three parameters. A <50% (from 3024×4032 to 1512×2016 pixels) reduction in image size was considered acceptable. Emboss filtering, in combination with MPII, provided the best results (reduced error of <60 pixels).

• Journal of the Korea Academia-Industrial cooperation Society
• /
• v.19 no.1
• /
• pp.672-679
• /
• 2018

This paper describes a methodology for shape design using an optimal design system, whereas generally a three dimensional analysis is required for such designs. An automatic finite element mesh generation technique, which is based on fuzzy knowledge processing and computational geometry techniques, is incorporated into the system, together with a commercial FE analysis code and a commercial solid modeler. Also, with the aid of multilayer neural networks, the present system allows us to automatically obtain a design window, in which a number of satisfactory design solutions exist in a multi-dimensional design parameter space. The developed optimal design system is successfully applied to evaluate the structures that are used. This study used a stress gauge to measure the maximum stress affecting the parts of the side housing bracket which are most vulnerable to cracking. Thereafter, we used a tool to interpret the maximum stress value, while maintaining the same stress as that exerted on the spot. Furthermore, a stress analysis was performed with the typical shape maintained intact, SM490 used for the material and the minimizing weight safety coefficient set to 3, while keeping the maximum stress the same as or smaller than the allowable stress. In this paper, a side housing bracket with a comparably simple structure for 36 tons was optimized, however if the method developed in this study were applied to side housing brackets of different classes (tons), their quality would be greatly improved.

• Journal of KIISE:Computing Practices and Letters
• /
• v.15 no.5
• /
• pp.345-349
• /
• 2009

Social bookmarking systems are a typical web 2.0 service based on folksonomy, providing the platform for storing and sharing bookmarking information. Spammers in social bookmarking systems denote the users who abuse the system for their own interests in an improper way. They can make the entire resources in social bookmarking systems useless by posting lots of wrong information. Hence, it is important to detect spammers as early as possible and protect social bookmarking systems from their attack. In this paper, we applied a diverse set of machine learning approaches, i.e., decision tables, decision trees (ID3), $na{\ddot{i}}ve$ Bayes classifiers, TAN (tree-augment $na{\ddot{i}}ve$ Bayes) classifiers, and artificial neural networks to this task. In our experiments, $na{\ddot{i}}ve$ Bayes classifiers performed significantly better than other methods with respect to the AUC (area under the ROC curve) score as veil as the model building time. Plausible explanations for this result are as follows. First, $na{\ddot{i}}ve$> Bayes classifiers art known to usually perform better than decision trees in terms of the AUC score. Second, the spammer detection problem in our experiments is likely to be linearly separable.

• Journal of Institute of Control, Robotics and Systems
• /
• v.16 no.12
• /
• pp.1150-1158
• /
• 2010

In this study, the Polynomial-based Radial Basis Function Neural Networks is proposed as one of the recognition part of overall face recognition system that consists of two parts such as the preprocessing part and recognition part. The design methodology and procedure of the proposed pRBFNNs are presented to obtain the solution to high-dimensional pattern recognition problem. First, in preprocessing part, we use a CCD camera to obtain a picture frame in real-time. By using histogram equalization method, we can partially enhance the distorted image influenced by natural as well as artificial illumination. We use an AdaBoost algorithm proposed by Viola and Jones, which is exploited for the detection of facial image area between face and non-facial image area. As the feature extraction algorithm, PCA method is used. In this study, the PCA method, which is a feature extraction algorithm, is used to carry out the dimension reduction of facial image area formed by high-dimensional information. Secondly, we use pRBFNNs to identify the ID by recognizing unique pattern of each person. The proposed pRBFNNs architecture consists of three functional modules such as the condition part, the conclusion part, and the inference part as fuzzy rules formed in 'If-then' format. In the condition part of fuzzy rules, input space is partitioned with Fuzzy C-Means clustering. In the conclusion part of rules, the connection weight of pRBFNNs is represented as three kinds of polynomials such as constant, linear, and quadratic. Coefficients of connection weight identified with back-propagation using gradient descent method. The output of pRBFNNs model is obtained by fuzzy inference method in the inference part of fuzzy rules. The essential design parameters (including learning rate, momentum coefficient and fuzzification coefficient) of the networks are optimized by means of the Particle Swarm Optimization. The proposed pRBFNNs are applied to real-time face recognition system and then demonstrated from the viewpoint of output performance and recognition rate.