• Journal of the Institute of Convergence Signal Processing
• /
• v.9 no.3
• /
• pp.201-207
• /
• 2008

This paper presents the implementation of a Tiling System about Preprocessing functions of JPEG 2000. The system covers the JPEG 2000 standard and is designed to determine the size of the image, to expand the image area and to split input image into several tiles. In order to split the input image with the progressive transmission into several tiles and transmit a tile of this image to others, this system store this image into Frame Memory. Therefore, this is designed as the Finite State Machine (FSM) to sequence through specific patterns of states in a predetermined sequential manner by using Verilog-HDL and be designed to handle a maximum 5M image. Moreover, for identifying image size for expansion, we propose several formula which are based on remainder after division (rem). we propose the true table which determines the size of the image input patterns by using results of these formula. Under the condition of TSMC 0.25um ASIC library, gate count is 18,725 and maximum data arrival time is 18.94 [ns].

• Journal of the Institute of Convergence Signal Processing
• /
• v.6 no.1
• /
• pp.8-14
• /
• 2005

The implementation of vision system using CCD camera which measures the earing rate of aluminium CAN is represented in this paper. In order to optimize the input image, the object of the input image is separated and the position of the image is calibrated. In the preprocessing, the definition of image is improved by the histogram equalization, and then the edges of the input image are detected by the Robert mask. The heights of the four ears and angles of the aluminium CAN are measured manually with the digital vernier calipers in industry. It takes 30 seconds to measure manually the height of one direction of the aluminium CAN at least three times. However, when the proposed system in this paper is applied, it takes 0.02 seconds only. In conclusion, the efficiency of the proposed system is higher than that of the system used in the industry.

• Journal of Environmental Science International
• /
• v.30 no.12
• /
• pp.1053-1065
• /
• 2021

In this study, the prediction technology of Hydrological Quantitative Precipitation Forecast (HQPF) was improved by optimizing the weather predictors used as input data for machine learning. Results comparison was conducted using bias and Root Mean Square Error (RMSE), which are predictive accuracy verification indicators, based on the heavy rain case on August 21, 2021. By comparing the rainfall simulated using the improved HQPF and the observed accumulated rainfall, it was revealed that all HQPFs (conventional HQPF and improved HQPF 1 and HQPF 2) showed a decrease in rainfall as the lead time increased for the entire grid region. Hence, the difference from the observed rainfall increased. In the accumulated rainfall evaluation due to the reduction of input factors, compared to the existing HQPF, improved HQPF 1 and 2 predicted a larger accumulated rainfall. Furthermore, HQPF 2 used the lowest number of input factors and simulated more accumulated rainfall than that projected by conventional HQPF and HQPF 1. By improving the performance of conventional machine learning despite using lesser variables, the preprocessing period and model execution time can be reduced, thereby contributing to model optimization. As an additional advanced method of HQPF 1 and 2 mentioned above, a simulated analysis of the Local ENsemble prediction System (LENS) ensemble member and low pressure, one of the observed meteorological factors, was analyzed. Based on the results of this study, if we select for the positively performing ensemble members based on the heavy rain characteristics of Korea or apply additional weights differently for each ensemble member, the prediction accuracy is expected to increase.

• International Journal of Naval Architecture and Ocean Engineering
• /
• v.13 no.1
• /
• pp.641-649
• /
• 2021

Fuel oil consumption (FOC) must be minimized to determine the economic route of a ship; hence, the ship power must be predicted prior to route planning. For this purpose, a numerical method using test results of a model has been widely used. However, predicting ship power using this method is challenging owing to the uncertainty of the model test. An onboard test should be conducted to solve this problem; however, it requires considerable resources and time. Therefore, in this study, a deep feed-forward neural network (DFN) is used to predict ship power using deep learning methods that involve data pattern recognition. To use data in the DFN, the input data and a label (output of prediction) should be configured. In this study, the input data are configured using ocean environmental data (wave height, wave period, wave direction, wind speed, wind direction, and sea surface temperature) and the ship's operational data (draft, speed, and heading). The ship power is selected as the label. In addition, various treatments have been used to improve the prediction accuracy. First, ocean environmental data related to wind and waves are preprocessed using values relative to the ship's velocity. Second, the structure of the DFN is changed based on the characteristics of the input data. Third, the prediction accuracy is analyzed using a combination comprising five hyperparameters (number of hidden layers, number of hidden nodes, learning rate, dropout, and gradient optimizer). Finally, k-means clustering is performed to analyze the effect of the sea state and ship operational status by categorizing it into several models. The performances of various prediction models are compared and analyzed using the DFN in this study.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.17 no.12
• /
• pp.3330-3344
• /
• 2023

This paper proposes an antenna performance prediction model in the autonomous driving radar manufacturing process. Our research work is based upon a challenge dataset, Driving Radar Manufacturing Process Dataset, and a typical AutoML machine learning workflow engine, Pycaret open-source Python library. Note that the dataset contains the total 70 data-items, out of which 54 used as input features and 16 used as output features, and the dataset is properly built into resolving the multi-output regression problem. During the data regression analysis and preprocessing phase, we identified several input features having similar correlations and so detached some of those input features, which may become a serious cause of the multicollinearity problem that affect the overall model performance. In the training phase, we train each of output-feature regression models by using the AutoML approach. Next, we selected the top 5 models showing the higher performances in the AutoML result reports and applied the ensemble method so as for the selected models' performances to be improved. In performing the experimental performance evaluation of the regression prediction model, we particularly used two metrics, MAE and RMSE, and the results of which were 0.6928 and 1.2065, respectively. Additionally, we carried out a series of experiments to verify the proposed model's performance by comparing with other existing models' performances. In conclusion, we enhance accuracy for safer autonomous vehicles, reduces manufacturing costs through AutoML-Pycaret and machine learning ensembled model, and prevents the production of faulty radar systems, conserving resources. Ultimately, the proposed model holds significant promise not only for antenna performance but also for improving manufacturing quality and advancing radar systems in autonomous vehicles.

• The KIPS Transactions:PartB
• /
• v.9B no.3
• /
• pp.375-382
• /
• 2002

Handwritten character recognition is classified into on-line handwritten character recognition and off-line handwritten character recognition. On-line handwritten character recognition has made a remarkable outcome compared to off-line hacdwritten character recognition. This method can acquire the dynamic written information such as the writing order and the position of a stroke by means of pen-based electronic input device such as a tablet board. On the contrary, Any dynamic information can not be acquired in off-line handwritten character recognition since there are extreme overlapping between consonants and vowels, and heavily noisy images between strokes, which change the recognition performance with the result of the preprocessing. This paper proposes a method that effectively extracts the stroke including dynamic information of characters for off-line Korean handwritten character recognition. First of all, this method makes improvement and binarization of input handwritten character image as preprocessing procedure using watershed algorithm. The next procedure is extraction of skeleton by using the transformed Lu and Wang's thinning: algorithm, and segment pixel array is extracted by abstracting the feature point of the characters. Then, the vectorization is executed with a maximum permission error method. In the case that a few strokes are bound in a segment, a segment pixel array is divided with two or more segment vectors. In order to reconstruct the extracted segment vector with a complete stroke, the directional component of the vector is mortified by using right-hand writing coordinate system. With combination of segment vectors which are adjacent and can be combined, the reconstruction of complete stroke is made out which is suitable for character recognition. As experimentation, it is verified that the proposed method is suitable for handwritten Korean character recognition.

• The KIPS Transactions:PartB
• /
• v.13B no.4 s.107
• /
• pp.353-360
• /
• 2006

This paper presents a robust algorithm for segmenting a vehicle license plate area from a road image. We consider the features of license plates in three aspects : 1) edges due to the characters in the plate, 2) colors in the plate, and 3) geometric properties of the plate. In the preprocessing step, we compute the thresholds based on each feature to decide whether a pixel is inside a plate or not. A statistical approach is applied to the sample images to compute the thresholds. For a given road image, our algorithm binarizes it by using the thresholds. Then, we select three candidate regions to be a plate by searching the binary image with a moving window. The plate area is selected among the candidates with simple heuristics. This algorithm robustly detects the plate against the transformation or the difference of color intensity of the plate in the input image. Moreover, the preprocessing step requires only a small number of sample images for the statistical processing. The experimental results show that the algorithm has 97.8% of successful segmentation of the plate from 228 input images. Our prototype implementation shows average processing time of 0.676 seconds per image for a set of $1280{\times}960$ images, executed on a 3GHz Pentium4 PC with 512M byte memory.

• Journal of Internet Computing and Services
• /
• v.25 no.4
• /
• pp.47-56
• /
• 2024

This paper proposes a method for fine-tuning BERT-based natural language processing models to perform sentiment analysis on Korean review data. By varying the input sequence length during this process and comparing the performance, we aim to explore the optimal performance according to the input sequence length. For this purpose, text review data collected from the clothing shopping platform M was utilized. Through web scraping, review data was collected. During the data preprocessing stage, positive and negative satisfaction scores were recalibrated to improve the accuracy of the analysis. Specifically, the GPT-4 API was used to reset the labels to reflect the actual sentiment of the review texts, and data imbalance issues were addressed by adjusting the data to 6:4 ratio. The reviews on the clothing shopping platform averaged about 12 tokens in length, and to provide the optimal model suitable for this, five BERT-based pre-trained models were used in the modeling stage, focusing on input sequence length and memory usage for performance comparison. The experimental results indicated that an input sequence length of 64 generally exhibited the most appropriate performance and memory usage. In particular, the KcELECTRA model showed optimal performance and memory usage at an input sequence length of 64, achieving higher than 92% accuracy and reliability in sentiment analysis of Korean review data. Furthermore, by utilizing BERTopic, we provide a Korean review sentiment analysis process that classifies new incoming review data by category and extracts sentiment scores for each category using the final constructed model.

• Journal of the Korean Institute of Intelligent Systems
• /
• v.10 no.4
• /
• pp.343-350
• /
• 2000

In this paper, we propose the Multi-FPNN(Fuzzy Polynomial Neural Networks) model based on FNN and PNN(Polyomial Neural Networks) for optimal system identifacation. Here FNN structure is designed using fuzzy input space divided by each separated input variable, and urilized both in order to get better output performace. Each node of PNN structure based on GMDH(Group Method of Data handing) method uses two types of high-order polynomials such as linearane and quadratic, and the input of that node uses three kinds of multi-variable inputs such as linear and quadratic, and the input of that node and Genetic Algorithms(GAs) to identify both the structure and the prepocessing of parameters of a Multi-FPNN model. Here, HCM clustering method, which is carried out for data preproessing of process system, is utilized to determine the structure method, which is carried out for data preprocessing of process system, is utilized to determance index with a weighting factor is used to according to the divisions of input-output space. A aggregate performance inddex with a wegihting factor is used to achieve a sound balance between approximation and generalization abilities of the model. According to the selection and adjustment of a weighting factor of this aggregate abjective function which it is acailable and effective to design to design and optimal Multi-FPNN model. The study is illustrated with the aid of two representative numerical examples and the aggregate performance index related to the approximation and generalization abilities of the model is evaluated and discussed.

• Journal of Korea Multimedia Society
• /
• v.13 no.2
• /
• pp.216-224
• /
• 2010

A RAM-based Neural Network(RNN) has the advantages of processing speed and hardware implementation. In spite of these advantages, it has a saturation problem, weakness of repeated learning and extract of a generalized pattern. To resolve these problems of RNN, the 3DNS model using cumulative multi discriminator was proposed. But that model does not solve the saturation problem yet. In this paper, we proposed a adaptive weight cumulative neural net(AWCNN) using the adaptive weight neuron (AWN) for solving the saturation problem. The proposed nets improved a recognition rate and the saturation problem of 3DNS. We experimented with the MNIST database of NIST without preprocessing. As a result of experimentations, the AWCNN was 1.5% higher than 3DNS in a recognition rate when all input patterns were used. The recognition rate using generalized patterns was similar to that using all input patterns.