• Journal of the Korea Computer Graphics Society
• /
• v.27 no.5
• /
• pp.45-54
• /
• 2021

Recently, deep learning-based denoising approaches have been actively studied. In particular, with the advances of blind denoising techniques, it become possible to train a deep learning-based denoising model only with noisy images in an image domain where it is impossible to obtain a clean image. We no longer require pairs of a clean image and a noisy image to obtain a restored clean image from the observation. However, it is difficult to recover the target using a deep learning-based denoising model trained by only noisy images if the distribution of the noisy image is far from the distribution of the clean image. To address this limitation, unpaired image denoising approaches have recently been studied that can learn the denoising model from unpaired data of the noisy image and the clean image. ISCL showed comparable performance close to that of supervised learning-based models based on pairs of clean and noisy images. In this study, we propose suitable normalization techniques for each purpose of architectures (e.g., generator, discriminator, and extractor) of ISCL. We demonstrate that the proposed method outperforms state-of-the-art unpaired image denoising approaches including ISCL.

• Proceedings of the KSRS Conference
• /
• 2008.03a
• /
• pp.195-200
• /
• 2008

The accuracy of classifying pixels in HIRIS images is usually degraded by noisy bands since noisy bands may deform the typical shape of spectral reflectance. Proposed in this paper is a statistical method for noisy band removal which mainly makes use of the correlation coefficients between bands. Considering each band as a random variable, the correlation coefficient measures the strength and direction of a linear relationship between two random variables. While the correlation between two signal bands is high, existence of a noisy band will produce a low correlation due to ill-correlativeness and undirectedness. The application of the correlation coefficient as a measure for detecting noisy bands is under a two-pass screening scheme. This method is independent of the prior knowledge of the sensor or the cause resulted in the noise. The classification in this experiment uses the unsupervised k-nearest neighbor algorithm in accordance with the well-accepted Euclidean distance measure and the spectral angle mapper measure. This paper also proposes a hierarchical combination of these measures for spectral matching. Finally, a separability assessment based on the between-class and within-class scatter matrices is followed to evaluate the performance.

• Journal of Information Processing Systems
• /
• v.17 no.5
• /
• pp.1004-1019
• /
• 2021

To solve the problems of the low image contrast, fuzzy edge details and edge details missing in noisy image fusion, this study proposes a noisy infrared and visible light image fusion algorithm based on non-subsample contourlet transform (NSCT) and an improved bilateral filter, which uses NSCT to decompose an image into a low-frequency component and high-frequency component. High-frequency noise and edge information are mainly distributed in the high-frequency component, and the improved bilateral filtering method is used to process the high-frequency component of two images, filtering the noise of the images and calculating the image detail of the infrared image's high-frequency component. It can extract the edge details of the infrared image and visible image as much as possible by superimposing the high-frequency component of infrared image and visible image. At the same time, edge information is enhanced and the visual effect is clearer. For the fusion rule of low-frequency coefficient, the local area standard variance coefficient method is adopted. At last, we decompose the high- and low-frequency coefficient to obtain the fusion image according to the inverse transformation of NSCT. The fusion results show that the edge, contour, texture and other details are maintained and enhanced while the noise is filtered, and the fusion image with a clear edge is obtained. The algorithm could better filter noise and obtain clear fused images in noisy infrared and visible light image fusion.

• International Journal of Contents
• /
• v.2 no.1
• /
• pp.5-8
• /
• 2006

This paper describes a new mechanism of detecting object of interest from a noisy image, without using any a-priori knowledge about the target. It employs a parallel set of filters inspired upon biological findings of mammalian vision. In our proposed system, several basic features are extracted directly from original input visual stimuli, and these features are integrated based on their local competitive relations and statistical information. Through integration process, unnecessary features for detecting the target are spontaneously decreased, while useful features are enhanced. Experiments have been performed on a set of computer generated and real images corrupted with noise.

• Journal of Korea Society of Digital Industry and Information Management
• /
• v.8 no.3
• /
• pp.41-48
• /
• 2012

In this paper, we propose a method of edge detection for noisy image. The proposed method uses a progressive filter for noise reduction and a Sobel operator for edge detection. The progressive filter combines a median filter and a modified rational filter. The proposed method for noise reduction adjusts rational filter direction according to an edge in the image which is obtained by median filtering. Our method effectively attenuates the noise while preserving the image details. Edge detection is performed by a Sobel operator. This operator can be implemented by integer operation and is therefore relatively fast. Our proposed method not only preserves edge, but also reduces noise in uniform region. Thus, edge detection is well performed. Our proposed method could improve results using further developed Sobel operator. Experimental results show that our proposed method has better edge detection with correct positions than those by existing median and rational filtering methods for noisy image.

• International Journal of Computer Science & Network Security
• /
• v.21 no.7
• /
• pp.182-190
• /
• 2021

Deep learning is an advanced technology for large-scale data analysis, with numerous promising cases like image processing, object detection and significantly more. It becomes customarily to use transfer learning and fine-tune a pre-trained CNN model for most image recognition tasks. Having people taking photos and tag themselves provides a valuable resource of in-data. However, these tags and labels might be noisy as people who annotate these images might not be experts. This paper aims to explore the impact of noisy labels on fine-tuning pre-trained CNN models. Such effect is measured on a food recognition task using Food101 as a benchmark. Four pre-trained CNN models are included in this study: InceptionV3, VGG19, MobileNetV2 and DenseNet121. Symmetric label noise will be added with different ratios. In all cases, models based on DenseNet121 outperformed the other models. When noisy labels were introduced to the data, the performance of all models degraded almost linearly with the amount of added noise.

• Journal of radiological science and technology
• /
• v.46 no.6
• /
• pp.475-484
• /
• 2023

This study aimed to compare filters for reducing speckle noise in ultrasound images using clinical liver images. We acquired the clinical liver ultrasound images, and noisy images were obtained by adding 0.01, 0.05, 0.10, and 0.50 intensity levels of speckle noise to the liver images. The Wiener filter, median modified Wiener filter, gamma filter, and Lee filter were designed for the noisy images by setting window sizes at 3×3, 5×5, and 7×7. The coefficient of variation (COV) and contrast to noise ratio (CNR) were calculated to evaluate noise reduction and various filters. Moreover, the filter with the highest image quality was selected and quantitatively compared to a noisy image. As a result, COV and CNR showed the noise improved result when the Lee filter was applied. Furthermore, the Lee filter image with a window size of 7×7 was noted to possess approximately a minimum of 1.28 to a maximum of 3.38 times better COV and a minimum of 2.18 to a maximum of 5.50 times better CNR than the noisy image. In conclusion, we confirmed that the Lee filter was effective in reducing speckle noise and proved that an appropriate window size needs to be set considering blurring.

• Journal of the Korea Society of Computer and Information
• /
• v.6 no.2
• /
• pp.41-47
• /
• 2001

In this paper, I propose a detection algorithm that can reliably separate moving objects from noisy background in the image sequence received from a camera at the fixed position. The proposed algorithm consists of four processes: generation of the difference image between the input image and the reference image. multilevel quantization of the difference image, and multistage merging in the quantized image, detection of the moving object using a back propagation in a neural network. The test results show that the proposed algorithm can detect moving objects very effectively in noisy environment.

• The Journal of Korean Institute of Communications and Information Sciences
• /
• v.21 no.10
• /
• pp.2552-2562
• /
• 1996

In this paper, we propose a segmentation algorithm that can reliably separate moving objects from noisy background in the image sequance received from a camera at the fixed position. The proposed algorithm consists of three processes:generation of the difference image between the input image and the reference image, multilevel quantization of the difference image, and multistagemerging in the quantized image. The quantization process requantizes the difference image based on the multiple threshold values determined bythe histogram analysis. The merging starts from the seed region which created by using the highest threshold value and ends when termination conditions are met. the proposed method has been tested with various real imge sequances containing intruders. The test results show that the proposed algorithm can detect moving objects like intruders very effectively in the noisy environment.

• Proceedings of the IEEK Conference
• /
• 2003.11a
• /
• pp.149-152
• /
• 2003

This paper presents a method to find motion vectors that are closer to true motion with noisy images for simultaneous noise reduction and do-interlacing. The proposed method requires four interlaced field images: one noisy field image and three field images from which noise is already removed. The validation of motion provides accurate motion vectors and allows us to utilize them even in very noisy environment. The validated motion vectors are first used for the noise reduction, buffered and used later for the noise reduction and de -interlacing.