• The Journal of the Acoustical Society of Korea
• /
• v.39 no.5
• /
• pp.441-446
• /
• 2020

In this paper, we propose the α-Feature Map Scaling (α-FMS) method which extends the FMS method that was designed to enhance the discriminative power of feature maps of deep neural networks in Speaker Verification (SV) systems. The FMS derives a scale vector from a feature map and then adds or multiplies them to the features, or sequentially apply both operations. However, the FMS method not only uses an identical scale vector for both addition and multiplication, but also has a limitation that it can only add a value between zero and one in case of addition. In this study, to overcome these limitations, we propose α-FMS to add a trainable parameter α to the feature map element-wise, and then multiply a scale vector. We compare the performance of the two methods: the one where α is a scalar, and the other where it is a vector. Both α-FMS methods are applied after each residual block of the deep neural network. The proposed system using the α-FMS methods are trained using the RawNet2 and tested using the VoxCeleb1 evaluation set. The result demonstrates an equal error rate of 2.47 % and 2.31 % for the two α-FMS methods respectively.

• The Korean Journal of Applied Statistics
• /
• v.34 no.6
• /
• pp.979-993
• /
• 2021

Acoustic scene classification (ASC) categorizes an audio file based on the environment in which it has been recorded. This has long been studied in the detection and classification of acoustic scenes and events (DCASE). In this study, we considered the problem that ASC faces in real-world applications that the model used should have low-complexity. We compared several models that apply light-weight techniques. First, a base CNN model was proposed using log mel-spectrogram, deltas, and delta-deltas features. Second, depthwise separable convolution, linear bottleneck inverted residual block was applied to the convolutional layer, and Quantization was applied to the models to develop a low-complexity model. The model considering low-complexity was similar or slightly inferior to the performance of the base model, but the model size was significantly reduced from 503 KB to 42.76 KB.

• Journal of Internet Computing and Services
• /
• v.20 no.6
• /
• pp.29-36
• /
• 2019

When dealing with outdoor images in a variety of computer vision applications, the presence of shadow degrades performance. In order to understand the information occluded by shadow, it is essential to remove the shadow. To solve this problem, in many studies, involves a two-step process of shadow detection and removal. However, the field of shadow detection based on CNN has greatly improved, but the field of shadow removal has been difficult because it needs to be restored after removing the shadow. In this paper, it is assumed that shadow is detected, and shadow-less image is generated by using original image and shadow mask. In previous methods, based on CGAN, the image created by the generator was learned from only the aspect of the image patch in the adversarial learning through the discriminator. In the contrast, we propose a novel method using a discriminator that judges both the whole image and the local patch at the same time. We not only use the residual generator to produce high quality images, but we also use joint loss, which combines reconstruction loss and GAN loss for training stability. To evaluate our approach, we used an ISTD datasets consisting of a single image. The images generated by our approach show sharp and restored detailed information compared to previous methods.

• Journal of Positioning, Navigation, and Timing
• /
• v.12 no.4
• /
• pp.399-408
• /
• 2023

In this paper, we propose a solution that enables continuous indoor/outdoor positioning of smartphone users through the integration of Pedestrian Dead Reckoning (PDR) and GPS/WiFi signals. Considering that accurate step detection affects the accuracy of PDR, we propose a Deep Neural Network (DNN)-based technology to distinguish between walking and non-walking signals such as walking in place. Furthermore, in order to integrate PDR with GPS and WiFi signals, a technique is used to select a proper measurement by distinguishing between indoor/outdoor environments based on GPS Dilution of Precision (DOP) information. In addition, we propose a technology to adaptively change the measurement error covariance matrix by detecting measurement outliers that mainly occur in the indoor/outdoor transition section through a residual-based χ² test. It is verified through experiments on a testbed that these technologies significantly improve the performance of PDR and PDR/GPS/WiFi fingerprinting-based integrated pedestrian navigation.

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2022.06a
• /
• pp.205-208
• /
• 2022

Multi-exposure high dynamic range (HDR) image reconstruction, the task of reconstructing an HDR image from multiple low dynamic range (LDR) images in a dynamic scene, often produces ghosting artifacts caused by camera motion and moving objects and also cannot deal with washed-out regions due to over or under-exposures. While there has been many deep-learning-based methods with motion estimation to alleviate these problems, they still have limitations for severely moving scenes. They also require large parameter counts, especially in the case of state-of-the-art methods that employ attention modules. To address these issues, we propose a frequency domain approach based on the idea that the transform domain coefficients inherently involve the global information from whole image pixels to cope with large motions. Specifically we adopt Residual Fast Fourier Transform (RFFT) blocks, which allows for global interactions of pixels. Moreover, we also employ Depthwise Overparametrized convolution (DO-conv) blocks, a convolution in which each input channel is convolved with its own 2D kernel, for faster convergence and performance gains. We call this LFFNet (Lightweight Frequency Fusion Network), and experiments on the benchmarks show reduced ghosting artifacts and improved performance up to 0.6dB tonemapped PSNR compared to recent state-of-the-art methods. Our architecture also requires fewer parameters and converges faster in training.

• Journal of Digital Policy
• /
• v.3 no.1
• /
• pp.27-32
• /
• 2024

In this paper, proposes a super resolution method that enhances the quality of results by refining texture features, contrasting each, and utilizing the results as weights. For the improvement of quality, a precise and clear restoration result in details such as boundary areas is crucial in super resolution, along with minimizing unnecessary artifacts like noise. The proposed method constructs a residual block structure with multiple paths and skip-connections for feature estimation in conventional Convolutional Neural Network (CNN)-based super resolution methods to enhance quality. Additional learning is performed for sharpened and blurred image results for further texture analysis. By contrasting each super resolution result and allocating weights through this process, the proposed method achieves improved quality in detailed and smoothed areas of the image. The experimental results of the proposed method, evaluated using the PSNR and SSIM values as quality metrics, show higher results compared to existing algorithms, confirming the enhancement in quality.

• Proceedings of the Korean Institute of Information and Commucation Sciences Conference
• /
• 2021.10a
• /
• pp.297-299
• /
• 2021

Predicting future productivity of tight oil is an important task for analyzing residual oil recovery and reservoir behavior. In general, productivity prediction is made using the decline curve analysis(DCA). In this study, we intend to propose an effective model for predicting future production using deep learning-based recurrent neural networks(RNN), LSTM, and GRU algorithms. As input variables, the main parameters are oil, gas, water, which are calculated during the production of tight oil, and the type curve calculated through various cluster analyzes. the output variable is the monthly oil production. Existing empirical models, the DCA and RNN models, were compared, and an optimal model was derived through hyperparameter tuning to improve the predictive performance of the model.

• The Journal of the Acoustical Society of Korea
• /
• v.40 no.3
• /
• pp.234-240
• /
• 2021

In this paper, the mask-based speech enhancement is improved for effective speech recognition in noise environments. In the mask-based speech enhancement, enhanced spectrum is obtained by multiplying the noisy speech spectrum by the mask. The VoiceFilter (VF) model is used as the mask estimation, and the Spectrogram Inpainting (SI) technique is used to remove residual noise of enhanced spectrum. In this paper, we propose a combined loss to further improve speech enhancement. In order to effectively remove the residual noise in the speech, the positive part of the Triplet loss is used with the component loss. For the experiment TIMIT database is re-constructed using NOISEX92 noise and background music samples with various Signal to Noise Ratio (SNR) conditions. Source to Distortion Ratio (SDR), Perceptual Evaluation of Speech Quality (PESQ), and Short-Time Objective Intelligibility (STOI) are used as the metrics of performance evaluation. When the VF was trained with the mean squared error and the SI model was trained with the combined loss, SDR, PESQ, and STOI were improved by 0.5, 0.06, and 0.002 respectively compared to the system trained only with the mean squared error.

• Geophysics and Geophysical Exploration
• /
• v.23 no.3
• /
• pp.117-130
• /
• 2020

Noises can distort acquired geophysical data, leading to their misinterpretation. Potential noises sources include anthropogenic activity, natural phenomena, and instrument noises. Conventional denoising methods such as wavelet transform and filtering techniques, are based on subjective human investigation, which is computationally inefficient and time-consuming. Recently, many researchers attempted to implement neural networks to efficiently remove noise from geophysical data. This study aims to review and analyze different types of neural networks, such as artificial neural networks, convolutional neural networks, autoencoders, residual networks, and wavelet neural networks, which are implemented to remove different types of noises including seismic, transient electromagnetic, ground-penetrating radar, and magnetotelluric surveys. The review analyzes and summarizes the key challenges in the removal of noise from geophysical data using neural network, while proposes and explains solutions to the challenges. The analysis support that the advancement in neural networks can be powerful denoising tools for geophysical data.

• Journal of Internet Computing and Services
• /
• v.25 no.1
• /
• pp.99-107
• /
• 2024

Along with economic growth and industrial development, there is an increasing demand for various electronic components and device production of semiconductor, SMT component, and electrical battery products. However, these products may contain foreign substances coming from manufacturing process such as iron, aluminum, plastic and so on, which could lead to serious problems or malfunctioning of the product, and fire on the electric vehicle. To solve these problems, it is necessary to determine whether there are foreign materials inside the product, and may tests have been done by means of non-destructive testing methodology such as ultrasound ot X-ray. Nevertheless, there are technical challenges and limitation in acquiring X-ray images and determining the presence of foreign materials. In particular Small-sized or low-density foreign materials may not be visible even when X-ray equipment is used, and noise can also make it difficult to detect foreign objects. Moreover, in order to meet the manufacturing speed requirement, the x-ray acquisition time should be reduced, which can result in the very low signal- to-noise ratio(SNR) lowering the foreign material detection accuracy. Therefore, in this paper, we propose a five-step approach to overcome the limitations of low resolution, which make it challenging to detect foreign substances. Firstly, global contrast of X-ray images are increased through histogram stretching methodology. Second, to strengthen the high frequency signal and local contrast, we applied local contrast enhancement technique. Third, to improve the edge clearness, Unsharp masking is applied to enhance edges, making objects more visible. Forth, the super-resolution method of the Residual Dense Block (RDB) is used for noise reduction and image enhancement. Last, the Yolov5 algorithm is employed to train and detect foreign objects after learning. Using the proposed method in this study, experimental results show an improvement of more than 10% in performance metrics such as precision compared to low-density images.