• International Journal of Computer Science & Network Security
• /
• v.21 no.8
• /
• pp.65-70
• /
• 2021

Waste accumulation is becoming a significant challenge in most urban areas and if it continues unchecked, is poised to have severe repercussions on our environment and health. The massive industrialisation in our cities has been followed by a commensurate waste creation that has become a bottleneck for even waste management systems. While recycling is a viable solution for waste management, it can be daunting to classify waste material for recycling accurately. In this study, transfer learning models were proposed to automatically classify wastes based on six materials (cardboard, glass, metal, paper, plastic, and trash). The tested pre-trained models were ResNet50, VGG16, InceptionV3, and Xception. Data augmentation was done using a Generative Adversarial Network (GAN) with various image generation percentages. It was found that models based on Xception and VGG16 were more robust. In contrast, models based on ResNet50 and InceptionV3 were sensitive to the added machine-generated images as the accuracy degrades significantly compared to training with no artificial data.

• Journal of Korea Multimedia Society
• /
• v.22 no.9
• /
• pp.1000-1010
• /
• 2019

Image semantic segmentation and dehazing are key tasks in the computer vision. In recent years, researches in both tasks have achieved substantial improvements in performance with the development of Convolutional Neural Network (CNN). However, most of the previous works for semantic segmentation assume the images are captured in clear weather and show degraded performance under hazy images with low contrast and faded color. Meanwhile, dehazing aims to recover clear image given observed hazy image, which is an ill-posed problem and can be alleviated with additional information about the image. In this work, we propose a deep multi-task network for simultaneous semantic segmentation and dehazing. The proposed network takes single haze image as input and predicts dense semantic segmentation map and clear image. The visual information getting refined during the dehazing process can help the recognition task of semantic segmentation. On the other hand, semantic features obtained during the semantic segmentation process can provide cues for color priors for objects, which can help dehazing process. Experimental results demonstrate the effectiveness of the proposed multi-task approach, showing improved performance compared to the separate networks.

• Korean Journal of Artificial Intelligence
• /
• v.11 no.1
• /
• pp.17-24
• /
• 2023

With the continuous growth in the amount of data collected and analyzed, deep learning has become increasingly popular for extracting meaningful insights from various fields. However, hardware limitations pose a challenge for achieving meaningful results with limited data. To address this challenge, this paper proposes an algorithm that leverages the characteristics of convolutional neural networks (CNNs) to reduce the size of image datasets by 20% through smoothing and shrinking the size of images using color elements. The proposed algorithm reduces the learning time and, as a result, the computational load on hardware. The experiments conducted in this study show that the proposed method achieves effective learning with similar or slightly higher accuracy than the original dataset while reducing computational and time costs. This color-centric dataset construction method using image smoothing techniques can lead to more efficient learning on CNNs. This method can be applied in various applications, such as image classification and recognition, and can contribute to more efficient and cost-effective deep learning. This paper presents a promising approach to reducing the computational load and time costs associated with deep learning and provides meaningful results with limited data, enabling them to apply deep learning to a broader range of applications.

• Journal of Biomedical Engineering Research
• /
• v.44 no.4
• /
• pp.293-301
• /
• 2023

Scoliosis is a three-dimensional deformation of the spine that is a deformity induced by physical or disease-related causes as the spine is rotated abnormally. Early detection has a significant influence on the possibility of nonsurgical treatment. To train a deep learning model with preprocessed images and to evaluate the results with and without data augmentation to enable the diagnosis of scoliosis based only on a chest X-ray image. The preprocessed images in which only the spine, rib contours, and some hard tissues were left from the original chest image, were used for learning along with the original images, and three CNN(Convolutional Neural Networks) models (VGG16, ResNet152, and EfficientNet) were selected to proceed with training. The results obtained by training with the preprocessed images showed a superior accuracy to those obtained by training with the original image. When the scoliosis image was added through data augmentation, the accuracy was further improved, ultimately achieving a classification accuracy of 93.56% with the ResNet152 model using test data. Through supplementation with future research, the method proposed herein is expected to allow the early diagnosis of scoliosis as well as cost reduction by reducing the burden of additional radiographic imaging for disease detection.

• Proceedings of the Korea Water Resources Association Conference
• /
• 2021.06a
• /
• pp.183-183
• /
• 2021

Deep learning methods and their application have become an essential part of prediction and modeling in water-related research areas, including hydrological processes, climate change, etc. It is known that application of deep learning leads to high availability of data sources in hydrology, which shows its usefulness in analysis of precipitation, runoff, groundwater level, evapotranspiration, and so on. However, there is still a limitation on microclimate analysis and prediction with deep learning methods because of deficiency of gauge-based data and shortcomings of existing technologies. In this study, a real-time rainfall prediction model was developed from a sky image data set with convolutional neural networks (CNNs). These daily image data were collected at Chung-Ang University and Korea University. For high accuracy of the proposed model, it considers data classification, image processing, ratio adjustment of no-rain data. Rainfall prediction data were compared with minutely rainfall data at rain gauge stations close to image sensors. It indicates that the proposed model could offer an interpolation of current rainfall observation system and have large potential to fill an observation gap. Information from small-scaled areas leads to advance in accurate weather forecasting and hydrological modeling at a micro scale.

• Computers and Concrete
• /
• v.32 no.5
• /
• pp.475-486
• /
• 2023

Nondestructive evaluation (NDE) is an important task of civil engineering structure monitoring and inspection, but minor damage such as small cracks in local structure is difficult to observe. If cracks continued expansion may cause partial or even overall damage to the structure. Therefore, monitoring and detecting the structure in the early stage of crack propagation is important. The crack detection technology based on machine vision has been widely studied, but there are still some problems such as bad recognition effect for small cracks. In this paper, we proposed a deep learning method based on sweep signals to evaluate concrete surface crack with a width less than 1 mm. Two convolutional neural networks (CNNs) are used to analyze the one-dimensional (1D) frequency sweep signal and the two-dimensional (2D) time-frequency image, respectively, and the probability value of average damage (ADPV) is proposed to evaluate the minor damage of structural. Finally, we use the standard deviation of energy ratio change (ERVSD) and infrared thermography (IRT) to compare with ADPV to verify the effectiveness of the method proposed in this paper. The experiment results show that the method proposed in this paper can effectively predict whether the concrete surface is damaged and the severity of damage.

• 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 the Korea Society of Computer and Information
• /
• v.27 no.3
• /
• pp.33-43
• /
• 2022

In this paper, we propose an ensemble model facilitated by multi-channel palm images with attention U-Net models and pretrained convolutional neural networks (CNNs) for establishing a contactless palm-based user identification system using conventional inexpensive camera sensors. Attention U-Net models are used to extract the areas of interest including hands (i.e., with fingers), palms (i.e., without fingers) and palm lines, which are combined to generate three channels being ped into the ensemble classifier. Then, the proposed palm information-based user identification system predicts the class using the classifier ensemble with three outperforming pre-trained CNN models. The proposed model demonstrates that the proposed model could achieve the classification accuracy, precision, recall, F1-score of 98.60%, 98.61%, 98.61%, 98.61% respectively, which indicate that the proposed model is effective even though we are using very cheap and inexpensive image sensors. We believe that in this COVID-19 pandemic circumstances, the proposed palm-based contactless user identification system can be an alternative, with high safety and reliability, compared with currently overwhelming contact-based systems.

• Journal of Information Processing Systems
• /
• v.14 no.1
• /
• pp.205-217
• /
• 2018

Deep learning using convolutional neural networks (CNNs) is being studied in various fields of image recognition and these studies show excellent performance. In this paper, we compare the performance of CNN architectures, KCR-AlexNet and KCR-GoogLeNet. The experimental data used in this paper is obtained from PHD08, a large-scale Korean character database. It has 2,187 samples of each Korean character with 2,350 Korean character classes for a total of 5,139,450 data samples. In the training results, KCR-AlexNet showed an accuracy of over 98% for the top-1 test and KCR-GoogLeNet showed an accuracy of over 99% for the top-1 test after the final training iteration. We made an additional Korean character dataset with fonts that were not in PHD08 to compare the classification success rate with commercial optical character recognition (OCR) programs and ensure the objectivity of the experiment. While the commercial OCR programs showed 66.95% to 83.16% classification success rates, KCR-AlexNet and KCR-GoogLeNet showed average classification success rates of 90.12% and 89.14%, respectively, which are higher than the commercial OCR programs' rates. Considering the time factor, KCR-AlexNet was faster than KCR-GoogLeNet when they were trained using PHD08; otherwise, KCR-GoogLeNet had a faster classification speed.

• Journal of Software Assessment and Valuation
• /
• v.17 no.2
• /
• pp.161-172
• /
• 2021

Neuromorphic architecture is drawing attention as a next-generation computing that supports artificial intelligence technology with low energy. However, FPGA embedded boards based on Neuromorphic architecturehave limited resources due to size and power. In this paper, we compared and evaluated the image reduction method using the interpolation method that rescales the size without considering the feature points and the DCT (Discrete Cosine Transform) method that preserves the feature points as much as possible based on energy. The scaled images were compared and analyzed for accuracy through CNN (Convolutional Neural Networks) in a PC environment and in the Nengo framework of an FPGA embedded board.. As a result of the experiment, DCT based classification showed about 1.9% higher performance than that of interpolation representation in both CNN and FPGA nengo environments. Based on the experimental results, when the DCT method is used in a limited resource environment such as an embedded board, a lot of resources are allocated to the expression of neurons used for classification, and the recognition rate is expected to increase.