• Journal of Broadcast Engineering
• /
• v.26 no.2
• /
• pp.125-131
• /
• 2021

In recent years, Convolutional Neural Networks (CNNs) have achieved outstanding performance in the fields of computer vision such as image classification, object detection, visual quality enhancement, etc. However, as huge amount of computation and memory are required in CNN models, there is a limitation in the application of CNN to low-power environments such as mobile or IoT devices. Therefore, the need for neural network compression to reduce the model size while keeping the task performance as much as possible has been emerging. In this paper, we propose a method to compress CNN models by combining matrix decomposition methods of LR (Low-Rank) approximation and CP (Canonical Polyadic) decomposition. Unlike conventional methods that apply one matrix decomposition method to CNN models, we selectively apply two decomposition methods depending on the layer types of CNN to enhance the compression performance. To evaluate the performance of the proposed method, we use the models for image classification such as VGG-16, RestNet50 and MobileNetV2 models. The experimental results show that the proposed method gives improved classification performance at the same range of 1.5 to 12.1 times compression ratio than the existing method that applies only the LR approximation.

• Korean Journal of Remote Sensing
• /
• v.38 no.6_4
• /
• pp.1871-1877
• /
• 2022

In the past, satellite development required enormous budget and time, so only some developed countries possessed satellites. However, with the recent emergence of low-budget satellites such as micro-satellites, many countries around the world are participating in satellite development. Low-orbit and geostationary-orbit satellites are used in various fields such as environment and weather monitoring, precise change detection, and disasters. Recently, it has been actively used for monitoring through deep learning-based object-of-interest detection. Until now, Korea has developed satellites for national demand according to the space development plan, and the satellite image obtained through this is used for various purpose in the public and private sectors. Interest in satellite image is continuously increasing in Korea, and various contests are being held to discover ideas for satellite image application and promote technology development. In this special issue, we would like to introduce the topics that participated in the recently held 2022 Satellite Information Application Contest and research on the processing and utilization of KOMPSAT image data.

• Journal of Intelligence and Information Systems
• /
• v.26 no.4
• /
• pp.1-25
• /
• 2020

Sentiment Analysis (SA) is a Natural Language Processing (NLP) task that analyzes the sentiments consumers or the public feel about an arbitrary object from written texts. Furthermore, Aspect-Based Sentiment Analysis (ABSA) is a fine-grained analysis of the sentiments towards each aspect of an object. Since having a more practical value in terms of business, ABSA is drawing attention from both academic and industrial organizations. When there is a review that says "The restaurant is expensive but the food is really fantastic", for example, the general SA evaluates the overall sentiment towards the 'restaurant' as 'positive', while ABSA identifies the restaurant's aspect 'price' as 'negative' and 'food' aspect as 'positive'. Thus, ABSA enables a more specific and effective marketing strategy. In order to perform ABSA, it is necessary to identify what are the aspect terms or aspect categories included in the text, and judge the sentiments towards them. Accordingly, there exist four main areas in ABSA; aspect term extraction, aspect category detection, Aspect Term Sentiment Classification (ATSC), and Aspect Category Sentiment Classification (ACSC). It is usually conducted by extracting aspect terms and then performing ATSC to analyze sentiments for the given aspect terms, or by extracting aspect categories and then performing ACSC to analyze sentiments for the given aspect category. Here, an aspect category is expressed in one or more aspect terms, or indirectly inferred by other words. In the preceding example sentence, 'price' and 'food' are both aspect categories, and the aspect category 'food' is expressed by the aspect term 'food' included in the review. If the review sentence includes 'pasta', 'steak', or 'grilled chicken special', these can all be aspect terms for the aspect category 'food'. As such, an aspect category referred to by one or more specific aspect terms is called an explicit aspect. On the other hand, the aspect category like 'price', which does not have any specific aspect terms but can be indirectly guessed with an emotional word 'expensive,' is called an implicit aspect. So far, the 'aspect category' has been used to avoid confusion about 'aspect term'. From now on, we will consider 'aspect category' and 'aspect' as the same concept and use the word 'aspect' more for convenience. And one thing to note is that ATSC analyzes the sentiment towards given aspect terms, so it deals only with explicit aspects, and ACSC treats not only explicit aspects but also implicit aspects. This study seeks to find answers to the following issues ignored in the previous studies when applying the BERT pre-trained language model to ACSC and derives superior ACSC models. First, is it more effective to reflect the output vector of tokens for aspect categories than to use only the final output vector of [CLS] token as a classification vector? Second, is there any performance difference between QA (Question Answering) and NLI (Natural Language Inference) types in the sentence-pair configuration of input data? Third, is there any performance difference according to the order of sentence including aspect category in the QA or NLI type sentence-pair configuration of input data? To achieve these research objectives, we implemented 12 ACSC models and conducted experiments on 4 English benchmark datasets. As a result, ACSC models that provide performance beyond the existing studies without expanding the training dataset were derived. In addition, it was found that it is more effective to reflect the output vector of the aspect category token than to use only the output vector for the [CLS] token as a classification vector. It was also found that QA type input generally provides better performance than NLI, and the order of the sentence with the aspect category in QA type is irrelevant with performance. There may be some differences depending on the characteristics of the dataset, but when using NLI type sentence-pair input, placing the sentence containing the aspect category second seems to provide better performance. The new methodology for designing the ACSC model used in this study could be similarly applied to other studies such as ATSC.

• Journal of the Korean Society of Surveying, Geodesy, Photogrammetry and Cartography
• /
• v.24 no.4
• /
• pp.327-334
• /
• 2006

The aerial photos need in a simple object such as cartography and ground cover classification and also in a social objects such as the city plan, environment, disaster, transportation etc. However, the shadow, which includes when taking the aerial photos, makes a trouble to interpret the ground information, and also users, who need the photos in their field tasks, have a restriction. Generally the shadow occurs by the building and surface topography, and the detail cause is by changing of the illumination in an area. For removing the shadow this study uses the single image and processes the image without the source of image and taking situation. Also, applying the entropy minimization method it generates the 1-D gray-scale invariant image for creating the shadow edge mask and using the Canny edge detection creates the shadow edge mask, and finally by filtering in Fourier frequency domain creates the intrinsic image which recovers the 3-D color information and removes the shadow.

• Journal of Drive and Control
• /
• v.20 no.2
• /
• pp.15-23
• /
• 2023

In this study, two-dimensional location of crops for auto weeding was detected using deep learning. To construct a dataset for soybean detection, an image-capturing system was developed using a mono camera and single-board computer and the system was mounted on a weeding robot to collect soybean images. A dataset was constructed by extracting RoI (region of interest) from the raw image and each sample was labeled with soybean and the background for classification learning. The deep learning model consisted of four convolutional layers and was trained with a weakly supervised learning method that can provide object localization only using image-level labeling. Localization of the soybean area can be visualized via CAM and the two-dimensional position of the soybean was estimated by clustering the pixels associated with the soybean area and transforming the pixel coordinates to world coordinates. The actual position, which is determined manually as pixel coordinates in the image was evaluated and performances were 6.6(X-axis), 5.1(Y-axis) and 1.2(X-axis), 2.2(Y-axis) for MSE and RMSE about world coordinates, respectively. From the results, we confirmed that the center position of the soybean area derived through deep learning was sufficient for use in automatic weeding systems.

• Proceedings of the Plant Resources Society of Korea Conference
• /
• 2019.10a
• /
• pp.87-87
• /
• 2019

The consumption of herbal medicine and related products (herbal products) have increased in South Korea. At the same time the quality, safety, and efficacy of herbal products is being raised. Currently, the herbal products are standardized and controlled according to the requirements of the Korean Pharmacopoeia, the National Institute of Health and the Ministry of Public Health and Social Affairs. The validation of herbal products and their medicinal component is important, since many of these herbal products are composed of two or more medicinal plants. However, there are no tools to support the validation process. Interest in deep learning has exploded over the past decade, for herbal medicine using algorithms to achieve herb recognition, symptom related target prediction, and drug repositioning have been reported. In this study, individual images of four herbs (Panax ginseng C.A. Meyer, Atractylodes macrocephala Koidz, Poria cocos Wolf, Glycyrrhiza uralensis Fischer), actually sold in the market, were achieved. Certain image preprocessing steps such as noise reduction and resize were formatted. After the features are optimized, we applied GoogLeNet_Inception v4 model for herb image recognition. Experimental results show that our method achieved test accuracy of 95%. However, there are two limitations in the current study. Firstly, due to the relatively small data collection (100 images), the training loss is much lower than validation loss which possess overfitting problem. Secondly, herbal products are mostly in a mixture, the applied method cannot be reliable to detect a single herb from a mixture. Thus, further large data collection and improved object detection is needed for better classification.

• Journal of the Korea Society of Computer and Information
• /
• v.29 no.1
• /
• pp.77-84
• /
• 2024

In this paper, we propose a method for accurately and rapidly detecting defects in wire harnesses by utilizing computer vision to calculate six crucial measurement values: the length of crimped terminals, the dimensions (width) of terminal ends, and the width of crimped sections (wire and core portions). We employ Harris corner detection to locate object positions from two types of data. Additionally, we generate reference points for extracting measurement values by utilizing features specific to each measurement area and exploiting the contrast in shading between the background and objects, thus reflecting the slope of each sample. Subsequently, we introduce a method using the Euclidean distance and correction coefficients to predict values, allowing for the prediction of measurements regardless of changes in the wire's position. We achieve high accuracy for each measurement type, 99.1%, 98.7%, 92.6%, 92.5%, 99.9%, and 99.7%, achieving outstanding overall average accuracy of 97% across all measurements. This inspection method not only addresses the limitations of conventional visual inspections but also yields excellent results with a small amount of data. Moreover, relying solely on image processing, it is expected to be more cost-effective and applicable with less data compared to deep learning methods.

• KIPS Transactions on Computer and Communication Systems
• /
• v.12 no.9
• /
• pp.273-282
• /
• 2023

Recently, various technologies such as logistics automation and port operations automation with ICT technology are being developed to build smart ports. However, there is a lack of technology development for port safety and safety accident prevention. This paper proposes an AI-based shipping container loading safety management system for the prevention of safety accidents at container loading fields in ports. The system consists of an AI-based shipping container safety accident risk classification and storage function and a real-time safety accident monitoring function. The system monitors the accident risk at the site in real-time and can prevent container collapse accidents. The proposed system is developed as a prototype, and the system is ecaluated by direct application in a port.

• Journal of Broadcast Engineering
• /
• v.18 no.4
• /
• pp.643-646
• /
• 2013

We present an improved AdaBoost algorithm to avoid overfitting phenomenon. AdaBoost is widely known as one of the best solutions for object detection. However, AdaBoost tends to be overfitting when a training dataset has noisy samples. To avoid the overfitting phenomenon of AdaBoost, the proposed method divides positive samples into K clusters using k-means algorithm, and then uses only one cluster to minimize the training error at each iteration of weak learning. Through this, excessive partitions of samples are prevented. Also, noisy samples are excluded for the training of weak learners so that the overfitting phenomenon is effectively reduced. In our experiment, the proposed method shows better classification and generalization ability than conventional boosting algorithms with various real world datasets.

• Journal of KIISE
• /
• v.45 no.1
• /
• pp.15-21
• /
• 2018

Low-Rank Representation (LRR) based methods are widely used in many practical applications, such as face clustering and object detection, because they can guarantee high prediction accuracy when used to constructing graphs in graph - based semi-supervised learning. However, in order to solve the LRR problem, it is necessary to perform singular value decomposition on the square matrix of the number of data points for each iteration of the algorithm; hence the calculation is inefficient. To solve this problem, we propose an improved and faster LRR method based on the recently published Fast LRR (FaLRR) and suggests ways to introduce and optimize additional constraints on the underlying optimization goals in order to address the fact that the FaLRR is fast but actually poor in classification problems. Our experiments confirm that the proposed method finds a better solution than LRR does. We also propose Fast MLRR (FaMLRR), which shows better results when the goal of minimizing is added.