News

Congratulations to Qingyang Zhou on successfully defending his dissertation today! His thesis, titled “Advanced Techniques for Point Cloud Quality Assessment, Surface Reconstruction, and Coding,” was reviewed by his committee—chaired by Jay Kuo, with Antonio Ortega as a member and Stefanos Nikolaidis serving as the outside examiner. The committee praised the rigor and excellence of Qingyang’s research. Here is a summary of his thesis:Point clouds (PCs) play a critical role in 3D vision and graphics applications such as AR/VR, digital preservation, and medical imaging. However, their large data volume, vulnerability to quality degradation, and the need for surface reconstruction from sparse, noisy points pose long-standing challenges. In this work, we address these issues through three research thrusts: compression, quality assessment, and surface reconstruction. First, we propose GPCGC, a low-complexity geometry compression method that leverages vector quantization and block-level rate-distortion modeling to achieve fast encoding/decoding with competitive performance. Second, we introduce GPQA, an interpretable, saliency-guided quality assessment framework that projects local 3D structures into 2D patches and uses a green machine learning model to predict perceptual quality under both full- and no-reference settings. Third, we develop two surface reconstruction methods: GPSR, an unsupervised diffusion-based approach, and LPSR, a supervised variant under the green learning paradigm. Together, these contributions advance the efficiency, interpretability, and robustness of the point cloud processing pipeline, enabling practical deployment in resource-constrained environments.He generously shared his experiences in the MCL Lab with us:My PhD journey at the University of Southern California’s Media Communications Lab (MCL) has been one of the most rewarding chapters of my life. Under the guidance of Professor C.-C. Jay Kuo, and with the generous support of MCL alumni and lab members, I not only deepened my technical knowledge but [...]

Single image super-resolution (SISR) is an intensively studied topic in image processing. It aims at recovering a high-resolution (HR) image from its low-resolution (LR) counterpart. SISR finds wide real-world applications such as remote sensing, medical imaging, and biometric identification. Besides, it attracts attention due to its connection with other tasks (e.g., image registration, compression, and synthesis). 

The main challenge of SISR is the ill-pose issue. We recently have been developing a solution by providing reasonable performance and effectively reduced complexity. We propose a green U-shape method to progressively enhance the LR images from global structure to local details with increasing spatial sizes and conditional residual estimation. 

Seismic waves are mechanical waves generated by earthquakes that travel through the Earth. Body waves consist of fast, compressional primary (P) waves and slower, shear secondary (S) waves. With large datasets of seismogram recordings, researchers train machine learning models to automatically pinpoint P‑ and S‑wave arrival times. This is essential for real‑time seismic monitoring and early warning systems.

Our Green Learning framework streamlines this process while boosting interpretability. We begin by slicing raw seismic recordings into overlapping three‑channel windows and assigning each a continuous pseudo‑label (ranging from 0 to 1) that reflects how accurately it is aligned to a P‑ or S‑wave onset. Treating these windows as 3‑channel images, we extract multi‑scale features via multiple Saab transform layers and select the most powerful features at each scale using Relevant Feature Test (RFT) modules. An XGBoost regressor then produces a continuous output signal, from which P‑ and S‑wave arrivals are simply recovered by peak detection. Compared to the SotA deep learning model EQTransformer, this model uses far fewer parameters,

Image-text retrieval is a fundamental task in image understanding. This task aims to retrieve the most relevant information from another modality based on the given image or text. Recent approaches focus on training large neural networks to bridge the gap between visual and textual domains. However, these models are computationally expensive and not explainable regarding how the data from different modalities are aligned. End-to-end optimized models, such as large neural networks, can only output the final results, making it difficult for humans to understand the reasoning behind the model’s predictions.

Hence, we propose a green learning solution, Green Multi-Modal Alignment (GMA), for computational efficiency and mathematical transparency. We reduce trainable parameters to 3% compared to fine-tuning the whole image and text encoders. The model is composed of three modules, including (1) Clustering, (2) Feature Selection, and (3) Alignment. The clustering process divides the whole dataset into subsets by choosing similar image and text pairs, reducing the training sample’s divergence. The second module, feature selection, reduces the feature dimension and mitigates the computational requirements. The importance of each feature can be interpreted as statistical evidence supporting our reasoning. The alignment is conducted by linear projection, which guarantees the inverse projection in both direction retrievals, namely image-to-text and tex-to-image retrievals.

Experimental results show that our model can outperform the SOTA retrieval models in text-to-image and image-to-text retrieval on the Flick30k and MS-COCO datasets. Besides, our alignment process can incorporate visual and text encoder models trained separately and generalize well to unseen image-text pairs.

Video Camouflaged Object Detection (VCOD) focuses on identifying and segmenting objectsconcealed within the background scenes. These camouflaged objects closely resemble theirsurroundings by mimicking similar color patterns and textures, which poses significant challengescompared to conventional detection tasks.To address this problem, we have proposed a motion-enhanced approach that progressivelyrefines the detection results with multi-resolution search and motion-guided boosting. The videoframe is first screened under image level, and the inter-frame motions and background models are then corrected, considering all the video sequences. This method provides stable performance under popular VCOD datasets.