News

Single Image Super-Resolution (SISR) is a classic problem in image processing that entails reconstructing a high-resolution (HR) image from a low-resolution (LR) input. Since the introduction of deep learning (DL) approaches, end-to-end methods have become the dominant paradigm in SISR. However, their performance gains typically come at the expense of increased model complexity and a substantial growth in parameter count.

To address these computational and interpretability challenges, we propose Green U-Shaped Learning (GUSL). This method departs from the black-box nature of deep neural networks by establishing a transparent, multi-stage pipeline. The framework utilizes residual correction across multiple resolution levels, mimicking a U-shaped structure where global structural information is captured at lower resolutions, and high-frequency local details are refined at higher resolutions. This progressive approach not only better manages the ill-posed nature of super-resolution through stage-wise regularization but also significantly reduces the parameter count and training footprint compared to conventional deep learning models.

Furthermore, to tackle the inherent spatial heterogeneity of the SISR task, where content varies significantly between smooth, homogeneous backgrounds (“Easy” samples) and complex, high-frequency textures (“Hard” samples), we adopt a divide-and-conquer strategy. By explicitly distinguishing between these two categories, we isolate the reconstruction tasks. This ensures that computational resources and modeling capacity are focused on regions with high residual errors, while simultaneously preventing the over-processing of simple areas.

As 2025 comes to a close, we look back with gratitude on a year marked by growth, achievement, and abundant blessings at MCL. This year, we bid a heartfelt farewell to our esteemed graduates, whose impactful research has left a lasting legacy as they embark on exciting new journeys. At the same time, we were delighted to welcome new members into the MCL family—bringing fresh perspectives, energy, and enthusiasm for innovative research.

Through dedication, collaboration, and resilience, our community reached significant milestones, including the publication of outstanding work in leading journals and conferences. These accomplishments reflect the collective passion, perseverance, and excellence of every member of MCL.

From all of us at MCL, we wish you a Merry Christmas filled with love, laughter, and cheer. May this festive season bring peace, happiness, and many moments to cherish.

Mouse brains are often used as model systems in medical studies, as they share many similarities with the human brain in function and structure. By observing biomarkers in brain tissue, researchers can gain insight into how neurons and blood vessels interact to support processes such as learning, memory, and sensory perception (especially in response to triggers of content/stress and so on).

Our studies on 3D microscopic images of mouse brains involve two biomarkers: lectin and cFOS. Lectin is a protein which binds specifically to certain sugar molecules on the surface of cells lining blood vessels, and can therefore be stained with a fluorescent dye to mark the network of blood vessels. cFOS labelling, on the other hand, marks neurons that were recently active, serving as an “activity map” of brain regions responding to external stimuli or behavioural tasks. The detailed information of both biomarkers is obtained through the dissection of the mouse (especially cFOS, which requires observation around 2 hours after stimulation), which means that it cannot be observed through imaging of the human brain.

These datasets are typically large per brain, high-dimensional, and require pixel-level interpretation. Manual labelling for such data is often time-consuming, and could vary depending on the experience level of the person labelling. A single brain can produce gigabytes to terabytes of 3D imaging data, containing complex vessel networks and cellular activation patterns. As cFOS shows up as very small points on the image, labelling those accurately proves to be difficult in practice as well.

Green learning offers an efficient method for automatically segmenting and analysing these complex images. Deep learning often demands extensive labelled data and computational resources, and such data is often hard to find when it comes to medical [...]

This work presents an interpretable and computationally efficient Video-Text Alignment (VTA) framework for cross-modal retrieval. Unlike end-to-end multimodal models that rely on large, opaque latent spaces and scale poorly with video length, VTA decomposes the retrieval process into transparent, modular components. The system first performs keyframe selection to reduce redundant temporal information and conducts object detection to extract salient visual concepts. In parallel, captions are analyzed through part-of-speech tagging to identify meaningful nouns and verbs. By modeling co-occurrence statistics between detected objects and textual keywords, VTA prunes irrelevant candidates early, dramatically reducing the search space and the number of trainable parameters—only 3% of those required for CLIP-based fine-tuning—while keeping encoders frozen to avoid overfitting to limited video-text datasets.

After filtering, VTA further improves retrieval through genre-based clustering and lightweight contrastive learning modules specialized to semantically coherent subsets of the data. These modules employ simple linear projections on top of frozen encoders, yet achieve competitive accuracy by focusing on more homogeneous sample groups. The entire pipeline is interpretable, as it provides explicit intermediate decisions, such as detected objects, POS-tagged keywords, genre predictions, and conditional probability estimates that link the two modalities. Experiments on the MSR-VTT benchmark demonstrate that VTA matches or surpasses state-of-the-art non-LLM baselines in both video-to-text and text-to-video retrieval, while offering constant-time inference and clear insight into its decision-making process.

Earthquakes generate seismic waves that travel through the Earth’s interior, carrying critical information about the Earth’s structure and the source event. These waves are mainly divided into two types: P-waves and S-waves. P-waves, or primary waves, travel fastest and arrive first at seismic stations, while S-waves, or secondary waves, follow with a slower speed but higher amplitude, often causing greater damage. Accurately identifying the arrival times of these waves — a process known as phase picking — is fundamental for earthquake localization, magnitude estimation, and early warning systems. However, seismic recordings are often contaminated by complex background noise, overlapping signals, and variable station conditions, which make manual picking both time-consuming and subjective. Automated phase picking, therefore, plays an increasingly vital role in modern seismology, enabling real-time earthquake detection and large-scale waveform analysis. Despite significant progress, achieving high precision and robustness across diverse seismic environments remains a major challenge for automated systems.

GreenPhase is a multi-resolution Green Learning framework for seismic phase picking. It aims to achieve high accuracy while maintaining interpretability and computational efficiency. The model operates across three resolution levels — from coarse to fine — progressively narrowing down candidate regions for P and S arrivals. At each level, GreenPhase extracts spectral–temporal features using the Saab transform and refines them through supervised feature selection and XGBoost regression. A pseudo-label generation and balanced sampling strategy further enhance training stability. Compared with deep networks such as PhaseNet and EQTransformer, GreenPhase requires far fewer training samples while achieving comparable performance. It is trained in a fully feedforward manner, balancing performance, efficiency, and interpretability. For the detection task, GreenPhase achieves an F1 score of 1.00; for P-phase picking, 0.98; and for S-phase picking, 0.96. The differences from EQTransformer, [...]