• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.25 no.11
• /
• pp.1627-1634
• /
• 2021

This paper introduces a low-power compact ADC circuit for analog Convolutional filter for low-power neural network accelerator SOC. While convolutional neural network accelerators can speed up the learning and inference process, they have drawback of consuming excessive power and occupying large chip area due to large number of multiply-and-accumulate operators when implemented in complex digital circuits. To overcome these drawbacks, we implemented an analog convolutional filter that consists of an analog multiply-and-accumulate arithmetic circuit along with an ADC. This paper is focused on the design optimization of a low-power 8bit SAR ADC for the analog convolutional filter accelerator We demonstrate how to minimize the capacitor-array DAC, an important component of SAR ADC, which is three times smaller than the conventional circuit. The proposed ADC has been fabricated in CMOS 65nm process. It achieves an overall size of 1355.7㎛², power consumption of 2.6㎼ at a frequency of 100MHz, SNDR of 44.19 dB, and ENOB of 7.04bit.

• Journal of IKEEE
• /
• v.27 no.3
• /
• pp.340-344
• /
• 2023

In this paper, we propose a study on the development of deep learning structure to secure visibility of outdoor LED display board according to weather change. The proposed technique secures the visibility of the outdoor LED display board by automatically adjusting the LED luminance according to the weather change using deep learning using an imaging device. In order to automatically adjust the LED luminance according to weather changes, a deep learning model that can classify the weather is created by learning it using a convolutional network after first going through a preprocessing process for the flattened background part image data. The applied deep learning network reduces the difference between the input value and the output value using the Residual learning function, inducing learning while taking the characteristics of the initial input value. Next, by using a controller that recognizes the weather and adjusts the luminance of the outdoor LED display board according to the weather change, the luminance is changed so that the luminance increases when the surrounding environment becomes bright, so that it can be seen clearly. In addition, when the surrounding environment becomes dark, the visibility is reduced due to scattering of light, so the brightness of the electronic display board is lowered so that it can be seen clearly. By applying the method proposed in this paper, the result of the certified measurement test of the luminance measurement according to the weather change of the LED sign board confirmed that the visibility of the outdoor LED sign board was secured according to the weather change.

• Journal of Broadcast Engineering
• /
• v.26 no.4
• /
• pp.429-440
• /
• 2021

With the recent development of deep convolutional neural network learning, deep learning techniques applied to single image super-resolution are showing good results. One of the existing deep learning-based super-resolution techniques is RDN(Residual Dense Network), in which the initial feature information is transmitted to the last layer using residual dense blocks, and subsequent layers are restored using input information of previous layers. However, if all hierarchical features are connected and learned and a large number of residual dense blocks are stacked, despite good performance, a large number of parameters and huge computational load are needed, so it takes a lot of time to learn a network and a slow processing speed, and it is not applicable to a mobile system. In this paper, we use the residual dense structure, which is a continuous memory structure that reuses previous information, and the residual dense channel attention block using the channel attention method that determines the importance according to the feature map of the image. We propose a method that can increase the depth to obtain a large receptive field and maintain a concise model at the same time. As a result of the experiment, the proposed network obtained PSNR as low as 0.205dB on average at 4× magnification compared to RDN, but about 1.8 times faster processing speed, about 10 times less number of parameters and about 1.74 times less computation.

• Journal of Korea Water Resources Association
• /
• v.53 no.12
• /
• pp.1159-1172
• /
• 2020

This study applied deep convolution neural network based on U-Net and SegNet using long period weather radar data to very short-term rainfall prediction. And the results were compared and evaluated with the translation model. For training and validation of deep neural network, Mt. Gwanak and Mt. Gwangdeoksan radar data were collected from 2010 to 2016 and converted to a gray-scale image file in an HDF5 format with a 1km spatial resolution. The deep neural network model was trained to predict precipitation after 10 minutes by using the four consecutive radar image data, and the recursive method of repeating forecasts was applied to carry out lead time 60 minutes with the pretrained deep neural network model. To evaluate the performance of deep neural network prediction model, 24 rain cases in 2017 were forecast for rainfall up to 60 minutes in advance. As a result of evaluating the predicted performance by calculating the mean absolute error (MAE) and critical success index (CSI) at the threshold of 0.1, 1, and 5 mm/hr, the deep neural network model showed better performance in the case of rainfall threshold of 0.1, 1 mm/hr in terms of MAE, and showed better performance than the translation model for lead time 50 minutes in terms of CSI. In particular, although the deep neural network prediction model performed generally better than the translation model for weak rainfall of 5 mm/hr or less, the deep neural network prediction model had limitations in predicting distinct precipitation characteristics of high intensity as a result of the evaluation of threshold of 5 mm/hr. The longer lead time, the spatial smoothness increase with lead time thereby reducing the accuracy of rainfall prediction The translation model turned out to be superior in predicting the exceedance of higher intensity thresholds (> 5 mm/hr) because it preserves distinct precipitation characteristics, but the rainfall position tends to shift incorrectly. This study are expected to be helpful for the improvement of radar rainfall prediction model using deep neural networks in the future. In addition, the massive weather radar data established in this study will be provided through open repositories for future use in subsequent studies.

• Journal of IKEEE
• /
• v.27 no.1
• /
• pp.116-121
• /
• 2023

In this paper, we develop a deep learning structure for a complex microbial incubator that applies deep learning prediction result information. The proposed complex microbial incubator consists of pre-processing of complex microbial data, conversion of complex microbial data structure, design of deep learning network, learning of the designed deep learning network, and GUI development applied to the prototype. In the complex microbial data preprocessing, one-hot encoding is performed on the amount of molasses, nutrients, plant extract, salt, etc. required for microbial culture, and the maximum-minimum normalization method for the pH concentration measured as a result of the culture and the number of microbial cells to preprocess the data. In the complex microbial data structure conversion, the preprocessed data is converted into a graph structure by connecting the water temperature and the number of microbial cells, and then expressed as an adjacency matrix and attribute information to be used as input data for a deep learning network. In deep learning network design, complex microbial data is learned by designing a graph convolutional network specialized for graph structures. The designed deep learning network uses a cosine loss function to proceed with learning in the direction of minimizing the error that occurs during learning. GUI development applied to the prototype shows the target pH concentration (3.8 or less) and the number of cells (10⁸ or more) of complex microorganisms in an order suitable for culturing according to the water temperature selected by the user. In order to evaluate the performance of the proposed microbial incubator, the results of experiments conducted by authorized testing institutes showed that the average pH was 3.7 and the number of cells of complex microorganisms was 1.7 × 10⁸. Therefore, the effectiveness of the deep learning structure for the complex microbial incubator applying the deep learning prediction result information proposed in this paper was proven.

• Journal of radiological science and technology
• /
• v.44 no.6
• /
• pp.655-661
• /
• 2021

Positron emission tomography (PET) images is affected by acquisition time, short acquisition times results in low gamma counts leading to degradation of image quality by statistical noise. Noise2Void(N2V) is self supervised denoising model that is convolutional neural network (CNN) based deep learning. The purpose of this study is to evaluate denoising performance of N2V for PET image with a short acquisition time. The phantom was scanned as a list mode for 10 min using Biograph mCT40 of PET/CT (Siemens Healthcare, Erlangen, Germany). We compared PET images using NEMA image-quality phantom for standard acquisition time (10 min), short acquisition time (2min) and simulated PET image (S2 min). To evaluate performance of N2V, the peak signal to noise ratio (PSNR), normalized root mean square error (NRMSE), structural similarity index (SSIM) and radio-activity recovery coefficient (RC) were used. The PSNR, NRMSE and SSIM for 2 min and S2 min PET images compared to 10min PET image were 30.983, 33.936, 9.954, 7.609 and 0.916, 0.934 respectively. The RC for spheres with S2 min PET image also met European Association of Nuclear Medicine Research Ltd. (EARL) FDG PET accreditation program. We confirmed generated S2 min PET image from N2V deep learning showed improvement results compared to 2 min PET image and The PET images on visual analysis were also comparable between 10 min and S2 min PET images. In conclusion, noisy PET image by means of short acquisition time using N2V denoising network model can be improved image quality without underestimation of radioactivity.

• Journal of the Korea Computer Graphics Society
• /
• v.24 no.2
• /
• pp.1-9
• /
• 2018

In this paper, we propose a method to restore corrupted black and white facial images to color. Previous studies have shown that when coloring damaged black and white photographs, such as old ID photographs, the area around the damaged area is often incorrectly colored. To solve this problem, this paper proposes a method of restoring the damaged area of input photo first and then performing colorization based on the result. The proposed method consists of two steps: BEGAN (Boundary Equivalent Generative Adversarial Networks) model based restoration and CNN (Convolutional Neural Network) based coloring. Our method uses the BEGAN model, which enables a clearer and higher resolution image restoration than the existing methods using the DCGAN (Deep Convolutional Generative Adversarial Networks) model for image restoration, and performs colorization based on the restored black and white image. Finally, we confirmed that the experimental results of various types of facial images and masks can show realistic color restoration results in many cases compared with the previous studies.

• Journal of Broadcast Engineering
• /
• v.24 no.5
• /
• pp.703-712
• /
• 2019

Single image Super-Resolution (SISR) aims to generate a visually pleasing high-resolution image from its degraded low-resolution measurement. In recent years, deep learning - based super - resolution methods have been actively researched and have shown more reliable and high performance. A typical method is WaveletSRNet, which restores high-resolution images through wavelet coefficient learning based on feature maps of images. However, there are two disadvantages in WaveletSRNet. One is a big processing time due to the complexity of the algorithm. The other is not to utilize feature maps efficiently when extracting input image's features. To improve this problems, we propose an efficient single image super resolution method, named RDB-WaveletSRNet. The proposed method uses the residual dense block to effectively extract low-resolution feature maps to improve single image super-resolution performance. We also adjust appropriated growth rates to solve complex computational problems. In addition, wavelet packet decomposition is used to obtain the wavelet coefficients according to the possibility of large scale ratio. In the experimental result on various images, we have proven that the proposed method has faster processing time and better image quality than the conventional methods. Experimental results have shown that the proposed method has better image quality by increasing 0.1813dB of PSNR and 1.17 times faster than the conventional method.

• Journal of the Korean Society of Radiology
• /
• v.14 no.7
• /
• pp.991-1001
• /
• 2020

Supervised deep learning technologies for improving the image quality of computed tomography (CT) need a lot of training data. When input images have different characteristics with training images, the technologies cause structural distortion in output images. In this study, an imaging model based on the deep reinforcement learning (DRL) was developed for overcoming the drawbacks of the supervised deep learning technologies and reducing noise in CT images. The DRL model was consisted of shared, value and policy networks, and the networks included convolutional layers, rectified linear unit (ReLU), dilation factors and gate rotation unit (GRU) in order to extract noise features from CT images and improve the performance of the DRL model. Also, the quality of the CT images obtained by using the DRL model was compared to that obtained by using the supervised deep learning model. The results showed that the image accuracy for the DRL model was higher than that for the supervised deep learning model, and the image noise for the DRL model was smaller than that for the supervised deep learning model. Also, the DRL model reduced the noise of the CT images, which had different characteristics with training images. Therefore, the DRL model is able to reduce image noise as well as maintain the structural information of CT images.

• Journal of the Korea Computer Graphics Society
• /
• v.28 no.1
• /
• pp.1-9
• /
• 2022

Non-small cell lung cancer (NSCLC) accounts for a high proportion of 85% among all lung cancer and has a significantly higher mortality rate (22.7%) compared to other cancers. Therefore, it is very important to predict the prognosis after surgery in patients with non-small cell lung cancer. In this study, the types of preoperative chest CT image patches for non-small cell lung cancer patients with tumor as a region of interest are diversified into five types according to tumor-related information, and performance of single classifier model, ensemble classifier model with soft-voting method, and ensemble classifier model using 3 input channels for combination of three different patches using pre-trained ResNet and EfficientNet CNN networks are analyzed through misclassification cases and Grad-CAM visualization. As a result of the experiment, the ResNet152 single model and the EfficientNet-b7 single model trained on the peritumoral patch showed accuracy of 87.93% and 81.03%, respectively. In addition, ResNet152 ensemble model using the image, peritumoral, and shape-focused intratumoral patches which were placed in each input channels showed stable performance with an accuracy of 87.93%. Also, EfficientNet-b7 ensemble classifier model with soft-voting method using the image and peritumoral patches showed accuracy of 84.48%.