Kaifeng Yang

Kaifeng Yang

Auto-generated profile, based on publicly available information

Kaifeng Yang

Camera Department, Mi Phone, Advanced Research Team, Xiaomi Corporation, Beijing, China

Research Focus & Metrics

Academic Career Length

5Years (2022-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
4 July 2022 in IEEE Transactions on Instrumentation and Measurement
Verified
Journal Article
Closed access
KYKaifeng Yang
Kaifeng Yang
YLYuliang Liu
Yuliang Liu
SZShiwen Zhang
Shiwen Zhang
JCJiajian Cao
Jiajian Cao
The surface defect detection plays an important role in industrial production and directly affects production efficiency and product quality. In this article, we focus on the surface defect detection of heat sink and propose a method based on a Ghost-SE light U-Net (GSLU-Net). The GSLU-Net is a novel combination of the lightweight convolution module, self-attention mechanism, and fully convolutional network (FCN). It has a symmetrical architecture inspired by the U-Net. By introducing the Ghost module, which can generate feature maps with cheap operations, we reduce the computation cost while maintaining high accuracy. The Ghost module uses an ordinary convolution with small kernels to obtain original feature maps and then a depthwise convolution to generate more feature maps. The squeeze-and-excitation (SE) block is also introduced to improve the representational power of the network and further improve accuracy. It can adaptively recalibrate channelwise feature responses at a slight additional computational cost. Fewer downsampling layers and more skip connections enable this network to retain more location information and have stronger detection capabilities for tiny targets than previous FCNs. The ablation experiments validate that our improvements have played their due role. Then, the comparative experiments demonstrate that our network outperforms the state-of-the-art FCNs. The GSLU-Net reaches an accuracy of 97.96% at a speed of 14.115 ms per image on the dataset of heat sink surface defect, achieving a balance of efficiency and accuracy. The GSLU-Net also outperforms other networks on open datasets, further demonstrating the superiority of the GSLU-Net.

7 June 2024 in IEEE Transactions on Instrumentation and Measurement
Verified
Journal Article
Closed access
BDBowei Dong
Bowei Dong
YLYuliang Liu
Yuliang Liu
KYKaifeng Yang
Kaifeng Yang
JCJiajian Cao
Jiajian Cao
In recent years, video-based physiological parameter measurement has gained significant attention because of its non-contact and low-cost characteristics. While supervised learning-based methods have achieved state-of-the-art performance, less attention has been paid to the reliability of the reference labels. In most studies, finger pulse oximeters are commonly used to record photoplethysmography (PPG) signals, serving as ground truth rather than utilizing real facial PPG signals. However, the appropriateness of this substitution for reference labels has been scarcely investigated. In this paper, we introduce a contact PPG acquisition system to acquire contact PPG signals of faces and fingers simultaneously and analyze their similarities and differences quantitatively. Experimental results reveal substantial disparities in morphology and phase between the PPG signal collected from the finger (Finger-PPG) and the PPG signal collected from the face (Facial-PPG). To eliminate the above differences, we propose a new framework: A Contrastive learning-based method is used to generate PPG signals (Contrast-PPG) to enhance Finger-PPG as reference labels. A 3DCNN-based network is used as a PPG estimator. This framework combines supervised and unsupervised methods to further improve performance. Extensive experiments are conducted proving that the corrected PPG signals are closer to the real facial PPG signal, effectively improving the performance and generalization ability of the model.