- Generative Adversarial Networks and Image Synthesis
- Advanced Image Processing Techniques
- Face recognition and analysis
- Advanced Image Fusion Techniques
- Image Enhancement Techniques
- Image and Signal Denoising Methods
- Image and Video Quality Assessment
- Industrial Vision Systems and Defect Detection
- Video Surveillance and Tracking Methods
- Face and Expression Recognition
- Thin-Film Transistor Technologies
- Organic Electronics and Photovoltaics
- Image Processing Techniques and Applications
- Advanced Neural Network Applications
- Advanced Image and Video Retrieval Techniques
- Human Pose and Action Recognition
- Visual Attention and Saliency Detection
- Advanced Vision and Imaging
- Remote-Sensing Image Classification
- Advanced Computing and Algorithms
- Emotion and Mood Recognition
- Semiconductor materials and devices
- Biometric Identification and Security
- Digital Media Forensic Detection
- Video Analysis and Summarization
National Taichung University of Science and Technology
2017-2024
National Yunlin University of Science and Technology
2023-2024
Novatek Microelectronics (Taiwan)
2023
AU Optronics (Taiwan)
2016-2018
Research Center for Information Technology Innovation, Academia Sinica
2015-2016
National Yang Ming Chiao Tung University
2014-2016
Academia Sinica
2016
University of Southern California
2014-2015
Center for Information Technology
2015
Arizona State University
2010
Image restoration is a challenging ill-posed problem which also has been long-standing issue. In the past few years, convolution neural networks (CNNs) almost dominated computer vision and had achieved considerable success in different levels of tasks including image restoration. However, recently Swin Transformer-based model shows impressive performance, even surpasses CNN-based methods to become state-of-the-art on high-level tasks. this paper, we proposed called SUNet uses Transformer...
Low-Light Image Enhancement is a computer vision task which intensifies the dark images to appropriate brightness. It can also be seen as an ill-posed problem in image restoration domain. With success of deep neural networks, convolutional networks surpass traditional algorithm-based methods and become mainstream area. To advance performance enhancement algorithms, we propose network (HWMNet) based on improved hierarchical model: M-Net+. Specifically, use half wavelet attention block M-Net+...
Clothing retrieval and clothing style recognition are important practical problems. They have drawn a lot of attention in recent years. However, the photos collected existing datasets mostly front- or near-front view. There no designed to study influences different viewing angles on performance. To address view-invariant problem properly, we construct challenge dataset, called Multi-View dataset. This dataset not only has four views for each item, but also provides 264 attributes describing...
We present a novel multistage learning system, called grouping estimation fusion (GEF), for human age via facial images. The GEF consists of three stages: (1) grouping; (2) within groups; and (3) decision final estimation. In the first stage, faces are classified into different groups, where each group has range. second methods adopted to extract global features from whole face local components (e.g., eyes, nose, mouth). Each or feature is individually utilized in group. Thus, several...
An ensemble method for full-reference image quality assessment (IQA) based on the parallel boosting (ParaBoost) idea is proposed in this paper. We first extract features from existing metrics and train them to form basic scorers (BIQSs). Then, we select additional address specific distortion types construct auxiliary (AIQSs). Both BIQSs AIQSs are trained small subsets of certain and, as a result, they weak performers with respect wide variety distortions. Finally, adopt ParaBoost framework,...
A no-reference (NR) learning-based approach to assess image quality is presented in this paper. The devised features are extracted from wide perceptual domains, including brightness, contrast, color, distortion, and texture. These used train a model (scorer) which can predict scores. scorer selection algorithms utilized help simplify the proposed system. In final stage, ensemble method combine prediction results selected scorers. Two multiple-scale versions of also along with single-scale...
This paper deals with the problem of clothing retrieval in a recommendation system. We develop hierarchical deep search framework to tackle this problem. use pre-trained network model that has learned rich mid-level visual representations module 1. Then, 2, we add latent layer and have neurons learn hashes-like while fine-tuning it on dataset. Finally, 3 achieves fast using hash codes via coarse-to-fine strategy. large dataset where 161,234 clothes images are collected labeled. Experiments...
Developing an automatic age estimation method towards human faces continues to possess important role in computer vision and pattern recognition. Many studies regarding facial mainly focus on two aspects: aging feature extraction classification/regression model learning. To set our work apart from existing approaches, we consider a different aspect -system structuring, which is, under constrained condition: given fixed type learning method, how design framework improve the performance based...
Although deep learning-based methods for semantic segmentation have achieved prominent performance in the general image domain, high-resolution remote sensing images remains highly challenging. One challenge is large size. High-resolution can very high spatial resolution, resulting with hundreds of millions pixels. This makes it difficult learning models to process efficiently, as they typically require amounts memory and computational resources. Another complexity objects scenes images....
In this paper, we focus on the creation of general purpose 2-D image quality databases. Although there are many them, they still lack some important characteristics, such as high-definition resolution, diversified source images, more commonly seen distortions, and a larger amount test (distorted) images. To tackle problem, create database, which has higher resolution than most addition, collect 250 images from 10 categories, far other existing Moreover, generate distortions to represent real...
Image inpainting is an important task in computer vision. As admirable methods are presented, the inpainted image getting closer to reality. However, result still not good enough reconstructed texture and structure based on human Although recent advances hardware have enabled development of larger more complex models, there a need for lightweight models that can be used by individuals small-sized institutions. Therefore, we propose model combines specialized transformer with traditional...
We have previously investigated the automatic current compliance property for indium tin oxide (ITO) resistance random access memory (RRAM). Traditionally, purpose of protecting RRAM, it is necessary to set equipment during and forming processes RRAM devices. ITO devices, however, an intrinsic capability limit their current. This letter examines this in depth by applying a varied stop-voltage measurement method, where different negative stop voltages were adopted manipulate oxygen ions....
Deep convolutional neural networks (DCNN) have demonstrated their potential to generate reasonable results in image inpainting. Some existing method uses convolution surrounding features, then passes features by fully connected layers, and finally predicts missing regions. Although the final result is semantically reasonable, some blurred situations generated because standard used, which conditioned on effective pixels substitute values masked holes. In this paper, we introduce dense blocks...
The recent super-resolution (SR) techniques are divided into two directions. One is to improve PSNR and the other visual quality. We believe improving quality more important practical than blindly PSNR. In this paper we employ a generative adversarial network (GAN) new perceptual loss function for photo-realistic single image (SISR). Our main contributions as follows: propose dense block which uses complex connections between each layer build powerful generator. Next, quality, found set of...
In this paper, we propose a new divide-and-conquer based method, called fusion of multiple binary age-grouping-estimation systems, for human facial age estimation. Under specific constraint, such as given feature or classification/regression what is the better framework estimation? First employ binary-grouping systems group classification. Each face image will be classified into one two groups. Within groups, models are trained to estimate ages faces their respectively. We also investigate...
Image restoration is a low-level vision task which to restore degraded images noise-free images. With the success of deep neural networks, convolutional networks surpass traditional methods and become mainstream in computer area. To advance performanceof denoising algorithms, we propose blind real image network (SRMNet) by employing hierarchical architecture improved from U-Net. Specifically, use selective kernel with residual block on structure called M-Net enrich multi-scale semantic...
Image restoration is a challenging and ill-posed problem which also has been long-standing issue. In this paper, we proposed multi-branch model inspired from the Human Visual System (i.e., Retinal Ganglion Cells) for image deraindrop. The experiments show that architecture, called CMFNet, state-of-the-art performance results. source code pretrained models are available at https://github.com/FanChiMao/CMFNet. And interactive demonstration of deraindrop can be accessed https://reurl.cc/dXaeNg.
In this paper, we present a structured fusion method for facial age group classification. To utilize the of shape features and surface features, introduced region certainty (ROC) to not only control classification accuracy feature based system but also reduce needs on system. first stage, design two which can be used classify frontal faces with high accuracies. second is adopted then selected by statistical method. The combined SVM classifier offer rates. With properly adjusting ROC single...
This paper presents a comparison study on subjective quality scores obtained by both single stimulus (without reference) and triple (with methods. The TID2013 database is reevaluated approach, which realized absolute category rating (ACR). And the mean opinion score (MOS) provided along with represents results from triple-stimulus pair (3-stimulus PC) method. In end, correlation coefficient hypothesis testing are used to determine if there significant difference between sets of scores....
In this paper, a deep Convolutional Neural Network CNN based system, called Depthwise Separable (DSCNN) fusion for human facial age estimation is presented. This system includes following four stages. the first stage, data augmentation procedure utilized to enrich dataset. second pre-trained model fine-tuned gender classification task. For third three newly designed DSCNN estimators are conduct gender-specific grouped images from previous stage. The architectures of these DSCNNs constructed...
Currently, neural network based defect detection systems usually need to collect a large number of samples for training, and it takes lot manpower mark labels clean the subsequent data. This is time-consuming process, makes whole system less effective. In this paper, method fabric surface proposed. By training positive samples, can learn through without collecting negative defective which greatly shortens landing time system. Our proposed achieve 99% accuracy.
In this work, we presented a novel clothing brand logo prediction method which is rooted on dense-block based deep convolutional neural network for detection and recognition. To learn networks deeper more accurately, adopted dense blocks into to make connections between layers shorter. our propose several structure designs improve recognition accuracy logos. We also built new large-scale (CBL) dataset with the attribute information facilitate task. reduce complexity proposed framework, two...
This paper proposed a new objective video quality metric for multimedia videos based on different perspective. We extended one existing image assessment to by considering temporal information and converted it into some compensation factor correct the score obtained in spatial domain. After experiments, we find out this does work well Laboratory Image Video Engineering (LIVE) Quality Database is also competitive with other state-of-the-art methods.
This work presents a novel facial makeup detection method, which includes four steps: entropy information computation, feature extraction, selection and classification. To carry out this objective, first all face images are subject to the computation. Once of faces obtained, extraction step is applied instead original images. The extracted features further processed reduce redundant on vector, done by procedure. A statistical analysis approach chosen realize purpose, aims lower dimension...