• Korean Journal of Remote Sensing
• /
• v.22 no.4
• /
• pp.235-242
• /
• 2006

This study compared between tropospheric column ozone by applying the SAM method to TOMS and OMI data for northern summer. Tropospheric ozone from the SAM represents a peak over the tropical Atlantic, where it is related with biomass burning. This feature is also seen in the distribution of the model and CO. Additionally, enhancement of the SAM ozone over the Middle East, and South and North America agrees well with the model and CO distribution. However, the SAM results show more ozone than the model results over the northern hemisphere, especially the ocean (e.g. the North Pacific and the North Atlantic). The tropospheric ozone distribution from OMI data shows more ozone than that from TOMS data. This can be caused by different viewing angle, sampling frequency, and a-priori ozone profiles between OMI and TOMS. The correlation between the SAM tropospheric ozone and CO is better than that between the model and CO in the tropics. However, that correlation is reversed in the mid-latitude.

• Journal of Korean Tunnelling and Underground Space Association
• /
• v.21 no.3
• /
• pp.419-432
• /
• 2019

Most of deep learning model training was proceeded by supervised learning, which is to train labeling data composed by inputs and corresponding outputs. Labeling data was directly generated manually, so labeling accuracy of data is relatively high. However, it requires heavy efforts in securing data because of cost and time. Additionally, the main goal of supervised learning is to improve detection performance for 'True Positive' data but not to reduce occurrence of 'False Positive' data. In this paper, the occurrence of unpredictable 'False Positive' appears by trained modes with labeling data and 'True Positive' data in monitoring of deep learning-based CCTV accident detection system, which is under operation at a tunnel monitoring center. Those types of 'False Positive' to 'fire' or 'person' objects were frequently taking place for lights of working vehicle, reflecting sunlight at tunnel entrance, long black feature which occurs to the part of lane or car, etc. To solve this problem, a deep learning model was developed by simultaneously training the 'False Positive' data generated in the field and the labeling data. As a result, in comparison with the model that was trained only by the existing labeling data, the re-inference performance with respect to the labeling data was improved. In addition, re-inference of the 'False Positive' data shows that the number of 'False Positive' for the persons were more reduced in case of training model including many 'False Positive' data. By training of the 'False Positive' data, the capability of field application of the deep learning model was improved automatically.

• Korean Journal of Radiology
• /
• v.22 no.8
• /
• pp.1323-1331
• /
• 2021

Objective: To identify the association between renal tumor complexity and pathologic renal sinus invasion (RSI) and evaluate the usefulness of computed tomography tumor features for predicting RSI in patients with renal cell carcinoma (RCC). Materials and Methods: This retrospective study included 276 consecutive patients who underwent radical nephrectomy for RCC with a size of ≤ 7 cm between January 2014 and October 2017. Tumor complexity and anatomical renal sinus involvement were evaluated using two standardized scoring systems: the radius (R), exophytic or endophytic (E), nearness to collecting system or sinus (N), anterior or posterior (A), and location relative to polar lines (RENAL) nephrometry and preoperative aspects and dimensions used for anatomical classification (PADUA) system. CT-based tumor features, including shape, enhancement pattern, margin at the interface of the renal sinus (smooth vs. non-smooth), and finger-like projection of the mass, were also assessed by two independent radiologists. Univariable and multivariable logistic regression analyses were performed to identify significant predictors of RSI. The positive predictive value, negative predictive value (NPV), accuracy of anatomical renal sinus involvement, and tumor features were evaluated. Results: Eighty-one of 276 patients (29.3%) demonstrated RSI. Among highly complex tumors (RENAL or PADUA score ≥ 10), the frequencies of RSI were 42.4% (39/92) and 38.0% (71/187) using RENAL and PADUA scores, respectively. Multivariable analysis showed that a non-smooth margin and the presence of a finger-like projection were significant predictors of RSI. Anatomical renal sinus involvement showed high NPVs (91.7% and 95.2%) but low accuracy (40.2% and 43.1%) for RSI, whereas the presence of a non-smooth margin or finger-like projection demonstrated comparably high NPVs (90.0% and 91.3% for both readers) and improved accuracy (67.0% and 73.9%, respectively). Conclusion: A non-smooth margin or the presence of a finger-like projection can be used as a preoperative CT-based tumor feature for predicting RSI in patients with RCC.

• Journal of Internet of Things and Convergence
• /
• v.10 no.2
• /
• pp.1-15
• /
• 2024

The construction of smart communities is a new method and important measure to ensure the security of residential areas. In order to solve the problem of low accuracy in face recognition caused by distorting facial features due to monitoring camera angles and other external factors, this paper proposes the following optimization strategies in designing a face recognition network: firstly, a global graph convolution module is designed to encode facial features as graph nodes, and a multi-scale feature enhancement residual module is designed to extract facial keypoint features in conjunction with the global graph convolution module. Secondly, after obtaining facial keypoints, they are constructed as a directed graph structure, and graph attention mechanisms are used to enhance the representation power of graph features. Finally, tensor computations are performed on the graph features of two faces, and the aggregated features are extracted and discriminated by a fully connected layer to determine whether the individuals' identities are the same. Through various experimental tests, the network designed in this paper achieves an AUC index of 85.65% for facial keypoint localization on the 300W public dataset and 88.92% on a self-built dataset. In terms of face recognition accuracy, the proposed network achieves an accuracy of 83.41% on the IBUG public dataset and 96.74% on a self-built dataset. Experimental results demonstrate that the network designed in this paper exhibits high detection and recognition accuracy for faces in surveillance videos.

• Journal of the Institute of Electronics Engineers of Korea SP
• /
• v.42 no.4 s.304
• /
• pp.1-12
• /
• 2005

Image registration is a process to establish the spatial correspondence between images of the same scene, which are acquired at different view points, at different times, or by different sensors. This paper presents a new algorithm for robust registration of the images acquired by multiple sensors having different modalities; the EO (electro-optic) and IR(infrared) ones in the paper. The two feature-based and intensity-based approaches are usually possible for image registration. In the former selection of accurate common features is crucial for high performance, but features in the EO image are often not the same as those in the R image. Hence, this approach is inadequate to register the E0/IR images. In the latter normalized mutual Information (nHr) has been widely used as a similarity measure due to its high accuracy and robustness, and NMI-based image registration methods assume that statistical correlation between two images should be global. Unfortunately, since we find out that EO and IR images don't often satisfy this assumption, registration accuracy is not high enough to apply to some applications. In this paper, we propose a two-stage NMI-based registration method based on the analysis of statistical correlation between E0/1R images. In the first stage, for robust registration, we propose two preprocessing schemes: extraction of statistically correlated regions (ESCR) and enhancement of statistical correlation by filtering (ESCF). For each image, ESCR automatically extracts the regions that are highly correlated to the corresponding regions in the other image. And ESCF adaptively filters out each image to enhance statistical correlation between them. In the second stage, two output images are registered by using NMI-based algorithm. The proposed method provides prospective results for various E0/1R sensor image pairs in terms of accuracy, robustness, and speed.

• Journal of Korean Society of Environmental Engineers
• /
• v.34 no.4
• /
• pp.223-231
• /
• 2012

The purpose of this study is to investigate basically the mechanism of heat transfer by the resolution of complex fluid flow inside a sophisticated designed screw dryer for the treatment of sewage sludge by using numerical analysis and experimental study. By doing this, the result was quite helpful to obtain the design criteria for enhancing drying efficiency, thereby achieving the optimal design of a multiple screw type dryer for treating inorganic and organic sludge wastes. One notable design feature of the dryer was to bypass a certain of fraction of the hot combustion gases into the bottom of the screw cylinder, by the fluid flow induction, across the delicately designed holes on the screw surface to agitate internally the sticky sludges. This offers many benefits not only in the enhancement of thermal efficiency even for the high viscosity material but also greater flexibility in the application of system design and operation. However, one careful precaution was made in operation in that when distributing the hot flue gas over the lump of sludge for internal agitation not to make any pore blocking and to avoid too much pressure drop caused by inertial resistance across the lump of sludge. The optimal retention time for rotating the screw at 1 rpm in order to treat 200 kg/hr of sewage sludge was determined empirically about 100 minutes. The corresponding optimal heat source was found to be 150,000 kcal/hr. A series of numerical calculation is performed to resolve flow characteristics in order to assist in the system design as function of important system and operational variables. The numerical calculation is successfully evaluated against experimental temperature profile and flow field characteristics. In general, the calculation results are physically reasonable and consistent in parametric study. In further studies, more quantitative data analyses such as pressure drop across the type and loading of drying sludge will be made for the system evaluation in experiment and calculation.

• Journal of Intelligence and Information Systems
• /
• v.22 no.2
• /
• pp.127-142
• /
• 2016

Deep learning model is a kind of neural networks that allows multiple hidden layers. There are various deep learning architectures such as convolutional neural networks, deep belief networks and recurrent neural networks. Those have been applied to fields like computer vision, automatic speech recognition, natural language processing, audio recognition and bioinformatics where they have been shown to produce state-of-the-art results on various tasks. Among those architectures, convolutional neural networks and recurrent neural networks are classified as the supervised learning model. And in recent years, those supervised learning models have gained more popularity than unsupervised learning models such as deep belief networks, because supervised learning models have shown fashionable applications in such fields mentioned above. Deep learning models can be trained with backpropagation algorithm. Backpropagation is an abbreviation for "backward propagation of errors" and a common method of training artificial neural networks used in conjunction with an optimization method such as gradient descent. The method calculates the gradient of an error function with respect to all the weights in the network. The gradient is fed to the optimization method which in turn uses it to update the weights, in an attempt to minimize the error function. Convolutional neural networks use a special architecture which is particularly well-adapted to classify images. Using this architecture makes convolutional networks fast to train. This, in turn, helps us train deep, muti-layer networks, which are very good at classifying images. These days, deep convolutional networks are used in most neural networks for image recognition. Convolutional neural networks use three basic ideas: local receptive fields, shared weights, and pooling. By local receptive fields, we mean that each neuron in the first(or any) hidden layer will be connected to a small region of the input(or previous layer's) neurons. Shared weights mean that we're going to use the same weights and bias for each of the local receptive field. This means that all the neurons in the hidden layer detect exactly the same feature, just at different locations in the input image. In addition to the convolutional layers just described, convolutional neural networks also contain pooling layers. Pooling layers are usually used immediately after convolutional layers. What the pooling layers do is to simplify the information in the output from the convolutional layer. Recent convolutional network architectures have 10 to 20 hidden layers and billions of connections between units. Training deep learning networks has taken weeks several years ago, but thanks to progress in GPU and algorithm enhancement, training time has reduced to several hours. Neural networks with time-varying behavior are known as recurrent neural networks or RNNs. A recurrent neural network is a class of artificial neural network where connections between units form a directed cycle. This creates an internal state of the network which allows it to exhibit dynamic temporal behavior. Unlike feedforward neural networks, RNNs can use their internal memory to process arbitrary sequences of inputs. Early RNN models turned out to be very difficult to train, harder even than deep feedforward networks. The reason is the unstable gradient problem such as vanishing gradient and exploding gradient. The gradient can get smaller and smaller as it is propagated back through layers. This makes learning in early layers extremely slow. The problem actually gets worse in RNNs, since gradients aren't just propagated backward through layers, they're propagated backward through time. If the network runs for a long time, that can make the gradient extremely unstable and hard to learn from. It has been possible to incorporate an idea known as long short-term memory units (LSTMs) into RNNs. LSTMs make it much easier to get good results when training RNNs, and many recent papers make use of LSTMs or related ideas.

• KDI Journal of Economic Policy
• /
• v.13 no.1
• /
• pp.151-166
• /
• 1991

In this paper, a simple game-theoretic entry deterrence model is developed that integrates both limit pricing and predatory pricing. While there have been extensive studies which have dealt with predation and limit pricing separately, no study so far has analyzed these closely related practices in a unified framework. Treating each practice as if it were an independent phenomenon is, of course, an analytical necessity to abstract from complex realities. However, welfare analysis based on such a model may give misleading policy implications. By analyzing limit and predatory pricing within a single framework, this paper attempts to shed some light on the effects of interactions between these two frequently cited tactics of entry deterrence. Another distinctive feature of the paper is that limit and predatory pricing emerge, in equilibrium, as rational, profit maximizing strategies in the model. Until recently, the only conclusion from formal analyses of predatory pricing was that predation is unlikely to take place if every economic agent is assumed to be rational. This conclusion rests upon the argument that predation is costly; that is, it inflicts more losses upon the predator than upon the rival producer, and, therefore, is unlikely to succeed in driving out the rival, who understands that the price cutting, if it ever takes place, must be temporary. Recently several attempts have been made to overcome this modelling difficulty by Kreps and Wilson, Milgram and Roberts, Benoit, Fudenberg and Tirole, and Roberts. With the exception of Roberts, however, these studies, though successful in preserving the rationality of players, still share one serious weakness in that they resort to ad hoc, external constraints in order to generate profit maximizing predation. The present paper uses a highly stylized model of Cournot duopoly and derives the equilibrium predatory strategy without invoking external constraints except the assumption of asymmetrically distributed information. The underlying intuition behind the model can be summarized as follows. Imagine a firm that is considering entry into a monopolist's market but is uncertain about the incumbent firm's cost structure. If the monopolist has low cost, the rival would rather not enter because it would be difficult to compete with an efficient, low-cost firm. If the monopolist has high costs, however, the rival will definitely enter the market because it can make positive profits. In this situation, if the incumbent firm unwittingly produces its monopoly output, the entrant can infer the nature of the monopolist's cost by observing the monopolist's price. Knowing this, the high cost monopolist increases its output level up to what would have been produced by a low cost firm in an effort to conceal its cost condition. This constitutes limit pricing. The same logic applies when there is a rival competitor in the market. Producing a high cost duopoly output is self-revealing and thus to be avoided. Therefore, the firm chooses to produce the low cost duopoly output, consequently inflicting losses to the entrant or rival producer, thus acting in a predatory manner. The policy implications of the analysis are rather mixed. Contrary to the widely accepted hypothesis that predation is, at best, a negative sum game, and thus, a strategy that is unlikely to be played from the outset, this paper concludes that predation can be real occurence by showing that it can arise as an effective profit maximizing strategy. This conclusion alone may imply that the government can play a role in increasing the consumer welfare, say, by banning predation or limit pricing. However, the problem is that it is rather difficult to ascribe any welfare losses to these kinds of entry deterring practices. This difficulty arises from the fact that if the same practices have been adopted by a low cost firm, they could not be called entry-deterring. Moreover, the high cost incumbent in the model is doing exactly what the low cost firm would have done to keep the market to itself. All in all, this paper suggests that a government injunction of limit and predatory pricing should be applied with great care, evaluating each case on its own basis. Hasty generalization may work to the detriment, rather than the enhancement of consumer welfare.