# Object-guided contrastive language-image pre-training for zero-shot target recognition

## TL;DR

This paper introduces OG-CLIP, a new model for target recognition that improves accuracy by focusing on key object regions and using military knowledge.

## Contribution

OG-CLIP introduces a novel framework with knowledge-driven augmentation, ROI modules, and adaptive MRL for better zero-shot target recognition.

## Key findings

- OG-CLIP achieves 84.28% mean accuracy, surpassing CLIP by 11.36 percentage points.
- The ROI module effectively suppresses background noise and focuses on discriminative regions.
- Experiments show OG-CLIP performs well on military and civilian targets.

## Abstract

Target recognition is critical for security systems, but traditional Visual-Language Models (VLMs) like CLIP suffer from limited training data semantics, poor background suppression, and inflexible multi-resolution features. To address these, we propose Object-Guide CLIP (OG-CLIP), integrating three core enhancements: Knowledge graph-driven data augmentation: A 5000-category military knowledge graph and 1M image-text pairs via multi-source acquisition and knowledge-infused prompts. Target-centered ROI module: Fuses SAM 2-generated masks with ViT features to focus on discriminative regions and suppress background noise. Adaptive MRL: Resolves traditional MRL’s rigid granularity via 128D–1024D continuous features, dynamic dimension weighting, and cross-granularity semantic alignment. Experiments on 99 target categories (military aircraft, warships, civilian targets) show OG-CLIP achieves 84.28% mean Accuracy (mAcc), 11.36 percentage points higher than baseline CLIP. Ablation confirms contributions of each component, and OG-CLIP excels in complex scenarios. The proposed framework offers a scalable and adaptable vision-language modeling approach for military recognition, with future work focusing on dataset expansion and model lightweight optimization.