Research map / Detection and segmentation / CNN features
Dual Information Purification for Lightweight SAR Object Detection
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 DIPKD, a feature-based knowledge distillation method for lightweight SAR object detectors that denoises teacher features via wavelet attention, decouples target/non-target knowledge, and adds reverse information transfer from student predictions.
Problem
Knowledge distillation for SAR object detection is harmed by speckle noise and background clutter, which put misleading information into teacher features and hinder transfer of target knowledge to lightweight students.
Approach
Combines three modules: Selective Noise Suppression (attention-guided wavelet denoising of FPN features), Knowledge Level Decoupling (separate distillation of target and non-target regions with masks and scale weights), and Reverse Information Transfer (student classification scores used to build a feature that constrains student intermediate features).
Key results
- On SSDD with Faster-RCNN ResNet18 student, DIPKD reaches 60.2 mAP, versus 57.4 for the undistilled student and 58.9 for the ResNet50 teacher.
- On HRSID with RetinaNet, the student reaches 50.6 mAP versus 44.8 undistilled, with a 51.4 mAP reported for SSDD RetinaNet.
- Ablation on SSDD shows RIT (+2.3 mAP), SNS (+2.0), and KLD (+0.8) individually; all three combined give +2.8 mAP over the student.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply DIPKD distillation | Faster-RCNN ResNet18 with ResNet50 teacher · SSDD | mAP: 57.4 -> 60.2 | Helped |
| Apply DIPKD distillation | Faster-RCNN ResNet18 with ResNet50 teacher · HRSID | mAP: 57.4 -> 58.8 | Helped |
| Apply DIPKD distillation | RetinaNet ResNet18 with ResNet50 teacher · SSDD | mAP: 44.9 -> 51.4 | Helped |
| Apply DIPKD distillation | RetinaNet ResNet18 with ResNet50 teacher · HRSID | mAP: 44.8 -> 50.6 | Helped |
| Add SNS, KLD, and RIT (ablation) | Faster-RCNN ResNet18 DIPKD · SSDD | mAP: 57.4 -> 60.2 | Helped |
| Add SNS (ablation) | Faster-RCNN ResNet18 DIPKD · SSDD | mAP: 57.4 -> 59.4 | Helped |
| Add KLD (ablation) | Faster-RCNN ResNet18 DIPKD · SSDD | mAP: 57.4 -> 58.2 | Helped |
| Add RIT (ablation) | Faster-RCNN ResNet18 DIPKD · SSDD | mAP: 57.4 -> 59.7 | Helped |
| Vary temperature τ from 0.25 to 1.25 (ablation) | Faster-RCNN ResNet50-ResNet18 DIPKD · SSDD | mAP: 59.9; 60.2; 60.1; 59.9; 59.8 | No clear effect |
| Vary mask rate from 0% to 100% (ablation) | Faster-RCNN ResNet50-ResNet18 DIPKD · SSDD | mAP: 60.0; 60.2; 60.1; 59.7; 58.1 | Mixed |
Surprising findings
- The student model surpasses the teacher on some metrics after distillation (e.g., 60.2 vs 58.9 mAP on SSDD Faster-RCNN).
Limitations
- Gains on HRSID with Faster-RCNN are small (58.8 vs 57.4 student and 58.9 teacher), and the paper's own claim of surpassing the teacher is only on some metrics.
- Loss weights are unnormalized and hand-tuned with very small values (e.g., 5e-5, 4.5e-7), with limited justification.
- Evaluated only on two SAR datasets and two detectors with a single ResNet50-to-ResNet18 teacher-student pair.
Experimental setup
Datasets: SSDD HRSID
Benchmarks: SSDD HRSID
Metrics: mAP AP50 AP75 APS APM APL mAR
Baselines: FKD FRS FGD MGD ARSD TWA IFKD Student without distillation
Builds on: FGD DeFeat Feature Pyramid Network Faster R-CNN RetinaNet
Closest papers by meaning
- SARDet-100K: Towards Open-Source Benchmark and ToolKit for Large-Scale SAR Object Detection (NeurIPS 2024 · similarity 0.87)
- Unleashing Channel Potential: Space-Frequency Selection Convolution for SAR Object Detection (CVPR 2024 · similarity 0.83)
- SCKD: Semi-Supervised Cross-Modality Knowledge Distillation for 4D Radar Object Detection (AAAI 2025 · similarity 0.82)
- RadarDistill: Boosting Radar-based Object Detection Performance via Knowledge Distillation from LiDAR Features (CVPR 2024 · similarity 0.82)
- CrossKD: Cross-Head Knowledge Distillation for Object Detection (CVPR 2024 · similarity 0.80)
- FUSAR-GPT: A Spatiotemporal Feature-Embedded and Two-Stage Decoupled Visual Language Model for SAR Imagery (CVPR 2026 · similarity 0.80)
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.