Digital Media Forensics

• Exposing DeepFakes: DeepFakes are AI-synthesized face images/videos. We develop detection methods of DeepFakes.
  Publications: [TMM23], [IJCAI23], [ICASSP23], [JCB22], [PRCV22], [PRL22], [ECCV22], [ICIP22C], [ICIP22A], [ICIP22B], [PR22], [MIPR22], [ACCESS22], [ICASSP22C], [BOOKCHAP20], [TCSVT21], [JCB21B], [ICIP21B], [ICIP21A], [IJCAI21], [SADFE21], [CVPR21], [ICASSP21], [BOOKCHAP21], [WIFS20], [INTERSPEECH20], [ICMEW20], [CVPR20], [CVPRW19B], [CVPRW19A], [IH19B], [ICASSP19A], [IH19A], [WIFS18], See also the Media Coverage of our work on DeepFake Forensics, and the DeepFake evaluation dataset Celeb-DF [link].
• Region Splicing Detection: Region splicing is a common image forgery operation. We developed several effective techniques for the detection of region splicing in images using blind estimation of image noise levels.
  Publications: [WIFS16], [IJCV14], [ICASSP12], [ICCP12], [MMSEC11]
• Region Copy-move Detection: Region copy-move is another common image forgery operation. we developed the first detection method based on the use of image features.
  Publications: [TIFS10], [ICASSP10B]
• Detecting Contrast Mapping: Contrast mapping is a very common image processing operation and can be used as an evidence for image tampering. We develop effective methods to estimate contrast mapping from images.
  Publications: [ICPR18A], [TOMM18], [ICIP14A]
• Audio Forensics: Audios are also subject to manipulations, such as splicing, duplication, and synthesis. We have developed a few methods for detecting tamperings in audio signals.
  Publications: [WMF23B], [ICASSP22A], [INTERSPEECH20], [CVPRW19A], [ICASSP12], [SPIE05]
• Estimating Lens Vignetting: Lens vignetting is an artifacts due to the curved surface of camera lenses. We developped a method that can estimate the vignetting effect for forensic analysis.
  Publications: [MMSEC10], [ICIP10]
• Art Forensics: We apply signal and statistical analysis to arts to find different styles in Renaissance paintings.
  Publications: [ARTMATH05], [PNAS04], [IFAR06]
• CG vs. Photo: Images can be generated using computer graphics software, and it has legal implications in the investigation of child pornography.
  Publications: [SACV03], [SP05]
• Steganalysis: Steganography is the technique used to hide information in innocuous media. The task of steganalysis is to detect the existance of stegaongraphy in digital images. This work is the receipient of the IEEE Signal Processing Best Paper Award.
  Publications: [SPIE05], [TIFS06], [SPIE04], [IH02]
• Counter-forensics: Counter forensics aims to remove or hide traces in tampered media to avoid detection. It is useful to identify vulnerabilities of the detection methods.
  Publications: [TCYB22], [MM21], [SCIS18]

Computer Vision

• Image Restoration/Enhancement: Images are often subject to quality degradation such as blurring and noise. Using statistical models of natural images, we have developed a series of methods to restore the degraded image to better qualities.
  Publications: [PR21], [ICCV15C], [CVPR15], [ICIP14], [ICIP14B], [CVPR14], [MANDM14B], [MANDM14A], [CVPR09], [CVPR08], [SPIE07], [PAMI09], [NIPS06]
• Object Tracking: Tracking multiple moving objects in videos is an important task in computer vision and is an important part of video analysis system. We have developed novel multi-object tracking methods based on hypergraphs.
  Publications: [CVPR23], [ICIP22D], [CVIU21], [CVIU22], [CVIU20], [AAAI19A], [AIC18B], [AIC18A], [TIP17], [AVSS17], [AICITY17A], [AICITY17B], [OT17], [ICIP2017], [AAAI17], [TCYB17], [IJCV16], [TIP16], [OTITR15], [OT15], [PAMI16]
• Object Recognition/Detection: Object recognition and detection are the most important and fundamental tasks in computer vision. We have developed several methods for object recognition/detection.
  Publications: [CVPRW22], [BMVC21B], [AVSS21B], [AVSS21A], [CVIU21], [CVIU22], [ACMM20], [ECCV20A], [IJCNN20], [CVIU20], [ACMMM19], [AAAI19B], [ICPR18B], [ACCV12], [CVPR05A], [CVPR05B]
• Semantic Segmentation: Semantic segmentation is also known as scene parsing, which aims to label every pixel in an image to sematincally meaningful labels.
  Publications: [TNNLS23], [JCB21A], [ICIP20], [AVSS19], [AVSS18]
• Image Annotation: Image annotation provides labels to objects in an image.
  Publications: [IJCV18], [AAAI16], [ICCV15B], [PR15]
• Natural Image Statistics: Natural images are not randomly sampled pixels, and there are regular statistical properties in them. We have developped several statistical models for natural images.
  Publications: [NIPS10], [NC11], [NC09], [MM05], [NIPS08]
• Human Pose Estimation: Human pose estimation is an important task in computer vision. It is the basis of human activity recognition and human centered robotic interaction.
  Publications: [TIP22], [AAAI22A], [ECCV20B], [TCYB21], [ECCV18], [ICIP18], [BMVC15], [FG15]
• Adversarial Perturbation: Adversarial perturbations are specially designed patterned noises that are added to images to disturb object detection algorithms.
  Publications: [ICME23B], [ICME23A], [ICME22], [ICCV21B], [ICCV21A], [ICIP21A], [ICME21], [IJCNN20], [BMVC19], [BMVC18]
• Human Activity Recognition: Analyzing human activities in videos is important for many applications in computer vision and video analysis.
  Publications: [IROS20], [AAAI20], [ICASSP19B], [ICCV17], [ICCV15A]
• 3D Photometric Reconstruction: We can reconstruct 3D structures from a collection of digital images. The
  Publications: [AVSS16], [ICIP15]
• Visual Saliency: Bottom up visual saliency is important to image retrieval and other computer vision tasks.
  Publications: [TCVG16], [SPIE15], [ICCE15]

Machine Learning

• Learning Objectives: Objective function is a universal and important component in most machine learning systems. We have studied new types of learning objectives.
  Publications: [PAMI23], [JMLR22], [NEURIPS20], [PAMI20], [ACCESS20], [NIPS17], [NIPS12], [NIPS11], [UAI09]
• AUROC Maximization: Area under ROC curves (AUROC) is an important performance metric for binary classification, we study direct optimization methods for AUROC.
  Publications: [AISTATS16], [FRONTIER19], [UAI18], [ICML18], [NIPS16], [OPT15]
• Deep Learning: Deep learning is to use deep neural networks to solve machine learning problems.
  Publications: [CVPR19], [IJCV19], [ACCV18], [CVPR18], [ICCV17], [IJCNN17], [ICCV15A], [ICANN14], [IJCAI13]
• Multi-label Learning: Multi-label learning is to assign an input data vector to one of many possible labels. Multi-label learning brings specific challenges to machine learning algorithms.
  Publications: [ACCESS20], [IJCV18], [AAAI16], [ICCV15B], [PR15]
• Matrix Factorization: Matrix factorization is an important machine learning problem, which aims to decompose a matrix into the product of several simpler matrices.
  Publications: [SPL19], [IJCAI19], [INFOSCI19], [PR18], [AAAI16], [ICLS14], [NIPSDDE13], [NIPS13], [ICASSP10A]
• Kernel Method: We designed new types of kernels for the task of object recognition.
  Publications: [CVPR05A], [ECML05]
• Boosting: We developed new boosting methods based on the infomax principle.
  Publications: [CVPR05B]

Robotics
In robotics, we developed new dynamic Bayesian inference algorithms for robotic perception.
Publications: [IROS23], [IROS18], [HUMANOIDS16], [IROS14], [ICRA15], [RSS14], [ICRA13]

Medical Imaging
We have developed several methods for the applications in medical imaging.
Publications: [ACCESS23], [FRONTIER23], [MIA21], [MICCAI19], [ISBI19], [SPIE12]