Scholarlink

Infrared-visible (IR-RGB) multimodal object detection leverages complementary information across modalities to enable robust all-weather object detection. However, existing methods often overlook the deep exploitation of modality-specific information and fail to model fine-grained local details, which are crucial for vehicle detection, leading to decreased robustness under complex lighting conditions. To address these challenges, we propose an Adaptive Fine-Grained Multimodal Feature Interaction Learning (AFMFIL) method, which integrates two key modules: Dynamic Frequency-Enhanced Local Interaction Learning (DFLIL) and Representative Negative Memory Contrastive Learning (RNMCL). The DFLIL module uses an adaptive dynamic frequency selection mechanism to extract key frequency information, promoting cross-modal feature interactions in local regions, thereby improving precise localization under complex lighting changes. Following this, the RNMCL module mines hard negative samples with high similarity across multiple feature levels to establish more comprehensive feature learning. This approach enables the model to learn more discriminative features among similar classes, thereby increasing inter-class distances and improving classification performance. Experimental results demonstrate that our method significantly improves detection accuracy and classification performance under challenging lighting conditions.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Adaptive Local Fine-Grained Feature Learning for Multimodal UAV-Based Vehicle Detection

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Jingjing Wan,
  • Meng Yang

摘要

Infrared-visible (IR-RGB) multimodal object detection leverages complementary information across modalities to enable robust all-weather object detection. However, existing methods often overlook the deep exploitation of modality-specific information and fail to model fine-grained local details, which are crucial for vehicle detection, leading to decreased robustness under complex lighting conditions. To address these challenges, we propose an Adaptive Fine-Grained Multimodal Feature Interaction Learning (AFMFIL) method, which integrates two key modules: Dynamic Frequency-Enhanced Local Interaction Learning (DFLIL) and Representative Negative Memory Contrastive Learning (RNMCL). The DFLIL module uses an adaptive dynamic frequency selection mechanism to extract key frequency information, promoting cross-modal feature interactions in local regions, thereby improving precise localization under complex lighting changes. Following this, the RNMCL module mines hard negative samples with high similarity across multiple feature levels to establish more comprehensive feature learning. This approach enables the model to learn more discriminative features among similar classes, thereby increasing inter-class distances and improving classification performance. Experimental results demonstrate that our method significantly improves detection accuracy and classification performance under challenging lighting conditions.