Research

< Real-time Video Understanding System based on multi-modal VLM Processing>

 - Abnormality detection: Unsupervised Video Anomaly Detection for known and unknown event classification

 - Computer-vision: Object classification and tracking with various AI system configuration

 - Image processing : Pre-processing to Improve VLM performance (Dehaze, Low-light HDR, Super-resolution)

 - Frame Selection: Key-frame sampling algorithm to optimize sequential image inputs to VLM for video understanding

 - Light-weight : Knowledge Distillation and Quantization aware Training for high speed AI processing

 - Multi-modal VLM : Image + refined additional information to improve VLM inference performance

 - VLM+RAG: VLM with RAG (Retrieval-Augmented-Generation) system for suggested Action to be followed

 < Object and behavior Classification> 

    - AI Algorithm for Abnormal behavior Recognition

   - Object detection and classification with  Serial/Parallel Ensemble structures

   - Seamless Object tracking algorithm with multi-feature extraction

   - De-Identification for Privacy protection in Video

 < Image processing for Object Classifcation robust to environment variation> 

    - Super-Resolution for small objects and  low resolution images                   

    - HDR & Contrast Enhancement for Low-light and IR camera                      

    - Dehazing for object visibility enhancement in adverse weather conditions    

    - Scene Analyzer to understand image visibility 

 < Video processing for consistent Object Classification >

    - Video stabilization compensated Object Classification                                 

    - Motion Analyzer based object classification and tracking 

 < Multi-modal AI & Light-weight AI >

    - Multi-domain Adaptation and Domain Transfer for AI classification

    - Knowledge distillation for Model Quantization and domain adaptation 

< Context and Layout aware Document AI >

    - Document layout Analysis : Document parsing algorithms to find page layout, table and cell detection

    - De-Identification: Personal Privacy Information Detection in text/scanned/captured documents by LLM and OCR

    - Document Classification : Confidentiality classification for complex documents with mixed contexts