Research map / Vision-language and multimodal / Cross-modal learning
DCEvo: Discriminative Cross-Dimensional Evolutionary Learning for Infrared and Visible Image Fusion
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DCEvo, a joint infrared-visible image fusion and detection/segmentation framework that uses an evolutionary algorithm to tune loss coefficients, plus a discriminative enhancer and cross-dimensional embedding block.
Problem
Infrared and visible image fusion is usually optimized separately from downstream perception tasks, so fused images give only marginal gains in detection/segmentation and task feedback does not improve fusion; hand-tuned loss weights trade off the two objectives poorly.
Approach
Treats joint fusion and detection training as a multi-objective problem in which a genetic algorithm (selection, crossover, mutation, roulette-wheel selection) searches loss coefficients every epoch, combined with a sigmoid-based Discriminative Enhancer in encoder/decoder and a Cross-Dimensional Embedding block that injects detection features into fusion features.
Key results
- DCEvo achieves MI 4.54, SSIM 0.51, VIF 0.87, Qabf 0.64 on M3FD, outperforming baselines on MI and SSIM there.
- On M3FD detection with YOLOv8s, DCEvo reaches 62.13 mAP, the highest among compared fusion methods (next best EMMA 61.60).
- On FMB segmentation with Segformer-b1, DCEvo reaches 52.52 mIoU versus 51.79 for the best baseline (MetaFusion).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DCEvo instead of the strongest listed baseline | IVIF methods · RoadScene | MI: 3.62 -> 3.69 | Helped |
| Use DCEvo instead of the strongest listed baseline | IVIF methods · TNO | MI: 3.79 -> 4.22 | Helped |
| Use DCEvo instead of the strongest listed baseline | IVIF methods · M3FD | Qabf: 0.68 -> 0.64 | Hurt |
| Use DCEvo instead of the strongest listed baseline | IVIF methods · M3FD | SSIM: 0.51 -> 0.51 | No clear effect |
| Add evolutionary hyperparameter search instead of experience-based coefficient settings (ablation) | DCEvo · M3FD | mAP: 61.70 -> 62.13 | Helped |
| Add evolutionary hyperparameter search instead of equal loss coefficients (ablation) | DCEvo · FMB | mIoU: 51.33 -> 52.52 | Helped |
| Add cooperative training of fusion and detection networks instead of separate optimization (ablation) | DCEvo · M3FD | mAP: 61.70 -> 62.13 | Helped |
| Add the discriminative enhancer (ablation) | DCEvo fusion architecture · M3FD | mAP: 61.23 -> 61.56 | Helped |
| Add the discriminative enhancer and projection-based enhancement (ablation) | DCEvo fusion architecture · FMB | mIoU: 50.23 -> 51.65 | Helped |
| Replace the base fusion model with concatenation-based feature embedding (ablation) | DCEvo with DE + DP · FMB | mIoU: 51.65 -> 51.78 | Helped |
| Use cross-attention embedding instead of the base fusion model (ablation) | DCEvo with DE + DP · FMB | mIoU: 51.65 -> 51.92 | Helped |
| Add the proposed cross- and self-attention feature embedding (ablation) | DCEvo with DE + DP · FMB | mIoU: 51.92 -> 52.52 | Helped |
Surprising findings
- Evolutionary search of loss coefficients gives only small gains over experience-based or equal coefficients (e.g., 62.13 vs 61.73 mAP), suggesting limited sensitivity to the tuning method.
Limitations
- Improvements over baselines in fusion metrics are mixed; on M3FD the method is not best on VIF and Qabf.
- Evaluation of downstream gains uses a single detector and single segmentation model with small reported margins.
Experimental setup
Datasets: MSRS M3FD RoadScene TNO FMB
Benchmarks: M3FD RoadScene TNO FMB
Metrics: MI SSIM VIF Qabf mAP mIoU
Baselines: YDTR SwinFusion ReCoNet TarDAL SegMiF DDFM LRRNet CDDFuse MetaFusion TIM EMMA Text-IF
Builds on: CDDFuse TarDAL SegMiF Text-IF
Closest papers by meaning
- Domain Adaptation Guided Infrared and Visible Image Fusion (AAAI 2026 · similarity 0.89)
- Revisiting Generative Infrared and Visible Image Fusion Based on Human Cognitive Laws (NeurIPS 2025 · similarity 0.85)
- Enhancing Infrared Vision: Progressive Prompt Fusion Network and Benchmark (NeurIPS 2025 · similarity 0.85)
- Dispel Darkness for Better Fusion: A Controllable Visual Enhancer based on Cross-modal Conditional Adversarial Learning (CVPR 2024 · similarity 0.85)
- Diff-2-in-1: Bridging Generation and Dense Perception with Diffusion Models (ICLR 2025 · similarity 0.84)
- Direct Discriminative Optimization: Your Likelihood-Based Visual Generative Model is Secretly a GAN Discriminator (ICML 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.