Yucong Zhao

Heilongjiang University of Science and Technology

Author Statistics

Papers

Citation

H-Index

i-10 index

Research Trends

Author Order

Document Type

Co-Authors

Jiquan Ma

Heilongjiang University of Science and Technology

Yuanchen Wang

Xi'an Jiaotong University

Xiaonan Zhu

Bank of New York Mellon

Ping Wang

Jilin University

Ping Wang

Xi'an Jiaotong University

Qianyu Zhou

Shanghai Jiao Tong University

Shumin Zhao

Heilongjiang University of Science and Technology

Ping Wang

Peking University

Cooperative Institutions

Heilongjiang University of Science and Technology

Xi'an Jiaotong University

Jiangnan University

Heilongjiang University

Jilin University

China University of Mining and Technology

China Academy of Space Technology

Nanjing University of Aeronautics and Astronautics

University of South Florida

Southwest Jiaotong University

Author Statistics

Papers

Citation

H-Index

i-10 index

Research Field

Video Super-Resolution using Multi-Frames Fusion and Perceptual Loss

Journal of Physics Conference Series (2019)

Xiaonan Zhu Yucong Zhao Ping Wang Yuanchen Wang Jiquan Ma

Abstract Video super-resolution (VSR) has become more and more important in vision perception. Currently, most of video super-resolution methods used optical flow estimation by explicit motion compensation. We proposed a new network using multi-frames as input without motion compensation. In order to improve the performance of the reconstructed high-resolution (HR) frame, perceptual loss function is used to evaluate the deviation between the prediction and ground truth. Perceptual loss characteristic is able to improve the quality of image reconstruction on the vision, but with a low Peak Signal to Noise Ratio (PSNR). Experimental results show that compared with other methods, our model can recover more details in high-resolution image, although it does not get higher PSNR value.

Ground truth

Optical Flow

10.1088/1742-6596/1345/2/022028

Cite

Citations (0)

Application of Wavelet Transform in Image Super-Resolution

Journal of Physics Conference Series (2019)

Yucong Zhao Xiaonan Zhu Ping Wang Yuanchen Wang Jiquan Ma

Abstract Most of learning-based super-resolution (SR) methods reconstructed high resolution (HR) image using a very deep convolutional neural network (CNN). CNN-based methods are used to process low resolution images, the predicted image is too smooth and some texture details are ignored. In order to improve the quality of the recovered high resolution image, wavelet transform is applied at the end of our model. Firstly, the prediction image is obtained through the CNN model; then, the predicted image and the ground truth image are processed into four frequency channels by wavelet transform; at last, four frequency channels and original image are considered when calculating loss. Compared with the original CNN model, the results of adding wavelet transform is obviously improved.

Top-hat transform

Stationary wavelet transform

Second-generation wavelet transform

10.1088/1742-6596/1345/2/022029

Cite

Citations (0)

Enhancement of Low-Light Image Based on Wavelet U-Net

Journal of Physics Conference Series (2019)

Yuanchen Wang Xiaonan Zhu Yucong Zhao Ping Wang Jiquan Ma

In computer vision, low-light image enhancement has always been a challenging task caused by more lower signal to noise ratio. Some methods have been proposed to enhance the low-light image using fully convolution network. Using u-net as backbone, we introduce wavelet transform to conduct down-sampling and up-sampling operations. In order to recover more details, perceptual loss has been used to optimize the network parameters. Experiments show that our model can get better performance than the existing methods. We find that wavelet transform effectively improve the quality of low-light image enhancement.

Convolution (computer science)

Peak signal-to-noise ratio

10.1088/1742-6596/1345/2/022030

Cite

Citations (4)

A Robust Facial Landmark Detection in Uncontrolled Natural Condition

Journal of Physics Conference Series (2020)

Ping Wang Qianyu Zhou Yucong Zhao Shumin Zhao Jiquan Ma

Abstract Facial landmark detection in uncontrolled natural condition is a challenging task due to illumination, occlusion, expression, posture, blur and other factors. In order to make a stable and accurate prediction, a novel data augmentation strategy based on random mask is used to obtain an efficient training data set. Our work is motivated by the ability of artist’s spatio imagination from part to whole. The artist can find the corresponding key points according to the incomplete image, so as to realize the whole from the part. In the experiment, the corresponding incomplete image is first generated based on the random mask and the complete image, and then the incomplete image is fed into the neural network to fit the mapping between the incomplete image and the facial landmarks. In order to evaluate the performance, a simple regression convolutional neural network based on the improved VGG16 [1] was trained, and the latest technical results were obtained. Furthermore, our method is tested over both complete and incomplete face images. Extensive experiments reveal that our model can work well on different human face data set.

Landmark

Data set

Training set

10.1088/1742-6596/1693/1/012202

Cite

Citations (3)