Projects

Generic Object Detection in Challenging Visual Environments

Target: Small, Occluded, and Rare Objects

This project focuses on improving detection reliability for objects that are small, heavily occluded, or belong to visually rare categories. Traditional detectors often struggle when objects occupy only a few pixels, appear partially hidden, or lack strong representation in the training distribution. The system enhances spatial reasoning for dense scenes, recovers obscured object cues using contextual priors, and stabilizes category confidence for infrequent visual classes, giving consistent performance across crowded environments, long-tail distributions, and real-world edge cases where conventional object detectors tend to fail.

Unified Document Parsing with Fine-Grained Element Detection

Target: Document Parsing

This project focuses on improving detection reliability in complex document images containing diverse and densely arranged visual components. The system enhances spatial reasoning for dense layouts, integrates contextual priors to improve structural coherence, and stabilizes category confidence for components that typically challenge OCR and downstream document understanding. Beyond conventional layout elements such as text blocks, tables, figures, and forms, the pipeline supports citation extraction, handwritten and digital signature detection, strikethrough identification and removal, stamp and watermark localization, and other fine-grained document objects crucial for legal, academic, and financial document processing.