News

Congratulations to Aolin Feng for passing his Qualifying Exam! His thesis proposal is titled “ Green Image Coding: Principle, Implementation, and Performance Evaluation.” His Qualifying Exam Committee members include Jay Kuo (Chair), Antonio Ortega, Bhaskar Krishnamachari, Feng Qian, and Shanghao Teng (Outside Member). Here is a summary of his thesis proposal:

Image compression has long been dominated by two paradigms: the hybrid coding framework that adopts prediction and transform followed by scalar quantization, and deep learning-based codecs that leverage end-to-end optimization. While hybrid codecs offer reliable rate–distortion (RD) performance, they face a bottleneck for RD gain when integrating additional coding tools. In contrast, deep learning-based codecs achieve superior RD performance through end-to-end optimization but often suffer from high computational cost, especially for the decoder side.

The thesis proposal introduces Green Image Coding (GIC), a novel framework aiming for lightweight, RD-efficient, and scalable image compression. Different from the hybrid and deep learning-based coding, GIC is built upon two key designs: multigrid representation and vector quantization (VQ). The multigrid representation decomposes images into hierarchical layers, redistributing energy and reducing intra-layer content diversity. Each layer is then encoded using VQ-based techniques. 

Regarding the two key designs, we contribute both theoretical foundations and practical solutions: multigrid rate control and a set of advanced VQ techniques. For the multigrid rate control, we develop a theory that reduces a high-dimensional optimization to equivalent sequential parameter decisions, with comprehensive experimental validation. The theoretical conclusion guides the design of our rate control strategy, which improves the scalability and balance between rate and distortion. For the advanced VQ techniques, we start with tree-structured vector quantization (TSVQ) to build multi-rate coding capability, and propose an RD-oriented VQ codebook construction method. We also propose the cascade VQ strategy to tackle the [...]

This work introduces Green IR Drop (GIRD), an energy-efficient and high-performance static IR-drop estimation method built on green learning principles. GIRD processes IC design inputs in three stages. First, the circuit netlist is transformed into multichannel maps, from which joint spatial–spectral representations are extracted using PixelHop. Next, discriminant features are identified through the Relevant Feature Test (RFT). Finally, these selected features are passed to an eXtreme Gradient Boosting (XGBoost) regressor. Both PixelHop and RFT belong to the family of green learning tools. Thanks to their lightweight design, GIRD achieves a low carbon footprint with significantly smaller model sizes and reduced computational complexity. Moreover, GIRD maintains strong performance even with limited training data. Experimental results on both synthetic and real-world circuits confirm its superior effectiveness. In terms of efficiency, GIRD’s model size and floating-point operation count (FLOPs) are only about 10⁻³ and 10⁻² of those required by deep learning methods, respectively.

Eosinophils are a type of white blood cell that can both protect the body and cause disease. While they are essential for fighting certain parasitic infections, they are also a primary cause of many allergic conditions when they don’t function correctly. It’s important to understand these cells because their buildup and activation are the main reasons for tissue damage in diseases like asthma and Eosinophilic Esophagitis (EoE).

Our current work explores the use of a dictionary learning pipeline to get unsupervised representations of whole-slide images of EoE. Unlike deep learning methods that require back-propagation to optimize millions of parameters to get data representations, our method represents the data in a self-organized way, with no back-propagation at all.

Compared to many other machine learning methods, our new pipeline provides a more transparent and explainable approach, especially for medical image analysis with smaller, specialized datasets.

The tumor microenvironment includes many types of cells around the tumor. Doctors often assess it using biomarkers like PD-L1 or CD68. To measure how many cells express these markers—called positive cells—we use image analysis and machine learning models to identify positive cells and compute their count and ratio. Usually, immunohistochemistry (IHC) images are used as the input of the model because they are relatively low-cost, while multiplex immunofluorescence (IF) images are used as ground truth due to their high accuracy. A major research goal is to predict positive cell locations from IHC images.

Our current work explores the use of the Green U-shaped Learning (GUSL) pipeline to align input IHC images with the ground truth IF images. GUSL is well-suited for this task because it enables pixel-wise prediction from coarse to fine resolution. It can detect positive cells at a coarse level and progressively refine predictions. GUSL has also shown strong performance in related tasks like kidney segmentation.

Another approach we explore is using GUSL to generate predictions on other medical images, such as H&E-stained images, DAPI, and LAP2. By producing segmentation results from H&E, DAPI, LAP2, and IHC inputs, and then taking a weighted average, we aim to further improve prediction accuracy.Currently, many machine learning methods have been developed to solve this problem, but they often suffer from high computational cost (FLOPs) and large model size. In addition, the limited size of medical datasets presents another major challenge. Green learning offers a promising solution to these issues and contributes to noninvasive biomarker prediction in this research field, helping reduce the need for expensive and labor-intensive staining methods.

Video object detection is a demanding computer vision topic that extends static image-based detection by introducing camera motion and temporal dynamics. This brings significant challenges, such as occlusion, scene blurriness, and dynamic shape changes caused by object and camera movement. Nevertheless, the temporal correlations between frames and the motion patterns of objects also provide rich and valuable information. The object detection and tracking over time enables machines to understand dynamic scenes and make informed decisions in complex environments. Nowadays, video object detection has become essential for many real-world applications, including autonomous driving, intelligent surveillance, human-computer interaction, and video content analysis.

Existing image-based detection models have achieved remarkable success, offering excellent accuracy and real-time detection capabilities in static scenarios. However, directly applying these models to video introduces several issues. Specifically, image-based models treat video frames independently, ignoring temporal relationships across frames, which often leads to unstable detection results in complex scenes and redundant computations for similar consecutive frames. Moreover, in real-world scenarios, videos are typically stored in compressed formats before being uploaded or transmitted. Fully decompression of the video further increases the computational overhead.

We propose Motion-Assisted YOLO (MA-YOLO), an efficient video object detection strategy that leverages the motion information naturally embedded in compressed video streams while utilizing existing image-based detection models to address the aforementioned challenges. Specifically, we adopt YOLO-X variants as our base detector for static images. Rather than performing detection on every video frame, we detect objects only on selected keyframes and propagate the predictions to estimate detection results for intermediate frames. The proposed framework consists of three modules: (1) keyframe selection and sparse inference, (2) motion vector extraction and pixel-wise assignment, and (3) motion-guided decision propagation. By incorporating the keyframe-based detection and motion [...]