Xiao-Qian Liu

Shandong University
1 previous affiliation
4Machine learning papers tracked by Lacuna
Is this you?

Research Overview

Xiao-Qian Liu focuses on computer vision and pattern recognition, with a primary emphasis on scene text recognition, unsupervised domain adaptation, and multilingual learning. Recent work addresses the challenges of domain shift and incremental learning in complex visual environments. The broader trajectory also intersects with emerging challenges in feature disentanglement and neural network generalization.

Research Trajectory

The documented research trajectory of Xiao-Qian Liu spans from 2023 to 2024, characterized by a focused effort to improve the robustness and adaptability of text recognition systems in complex, real-world environments. A central theme of this early work is overcoming the limitations of domain drift—the degradation in model performance when deployed on visual data that differs from its training distribution.

In 2023, Liu's research primarily addressed cross-domain text recognition through unsupervised domain adaptation (UDA). As first author on Unsupervised Domain Adaptation via Class Aggregation for Text Recognition, Liu proposed a dual intra-Class Aggregation based unsupervised Domain Adaptation method (CADA). This work identified a critical gap in existing UDA approaches, which typically relied on global image alignment or broad semantic features. CADA advanced the field by aligning both visual and semantic character feature distributions, leveraging the stroke features of individual characters to improve domain transfer without requiring target-domain labels.

Continuing this line of inquiry, Liu contributed as second author to DOC: Text Recognition via Dual Adaptation and Clustering (2023). This research further tackled the severe domain shift problem in text image recognition by combining local-level (character) and global-level feature adaptation. By addressing the lack of character-level supervision in target domains, the work provided a more granular approach to feature alignment, ensuring that the categories of local-level characters were preserved during the adaptation process.

By 2024, Liu's focus expanded to encompass multilingual capabilities and the architectural bottlenecks of vision-language models. Leading the effort as first author on Hierarchical Multi-label Learning for Incremental Multilingual Text Recognition, Liu addressed the challenge of catastrophic forgetting in multilingual text recognition (MLTR). The paper identified the "incremental sharing problem"—the failure of standard incremental learning methods to account for shared words and characters across different languages. To solve this, Liu introduced the HAMMER framework, which utilizes online knowledge analysis to identify shared linguistic features and provide multi-label language supervision, allowing models to learn new languages while retaining previous knowledge.

In parallel, Liu contributed as second author to VOLTER: Visual Collaboration and Dual-Stream Fusion for Scene Text Recognition (2024). This work investigated the architectural limitations of mainstream scene text recognition models, which typically rely on language models to iteratively refine the predictions of a vision model. The research demonstrated that vision models often act as a performance bottleneck due to their inability to capture internal image dependencies at multiple scales. VOLTER addressed this by introducing visual collaboration and dual-stream fusion, effectively bridging the gap between visual and linguistic feature spaces.

Current Directions

The most recent papers in the evidence from 2024 indicate an ongoing focus on incremental learning frameworks and multi-scale visual processing for scene text recognition. However, the underlying methodologies of Liu's work—specifically domain adaptation, feature alignment, and the prevention of model overfitting to specific training distributions—connect to broader challenges in computer vision security and generalization.

Beyond explicit text recognition tasks, Liu's profile intersects with methodological themes concerning how neural networks generalize across shifting distributions and how specific artifacts can be isolated. This is reflected in connections to the research direction of Disentangling Content and Artifacts for Generalizable Deepfake Detection and the area of Generalizing Media Forensics to Future Generative Architectures. The core technical challenge in these areas—isolating robust, generalizable features rather than relying on domain-specific shortcuts—mirrors the domain adaptation strategies employed in Liu's text recognition research.

Furthermore, the broader area of Universal Forensic Signatures of Generative Neural Networks and the research direction of Tracing Synthetic Content to Specific Generative Models highlight an overarching thematic interest in the structural properties of generative architectures. While Liu's primary applied domain remains text recognition, the techniques used to align features and disentangle visual semantics are highly relevant to these emerging challenges in media forensics and model attribution.

Profile updated

See something wrong? Report a mistake