Research map / Detection and segmentation / CNN features
Unleashing Channel Potential: Space-Frequency Selection Convolution for 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 SFS-Conv, a convolution module that shunts channels into spatial (multi-scale residual kernels) and frequency (fractional Gabor) branches and fuses them parameter-free, and builds the lightweight SFS-CNet detector for SAR object detection.
Problem
Standard convolutions in DCNN-based SAR object detectors extract redundant features and are computationally expensive, while existing lightweight or compression methods degrade detection accuracy.
Approach
A shunt-perceive-select convolution that splits input channels into a spatial branch (multi-scale hierarchical-residual kernels with expanding receptive fields) and a frequency branch (convolution kernels modulated by fractional Gabor filters), then fuses them via a parameter-free channel-wise soft-attention selection, increasing feature diversity within a single layer without adding channels.
Key results
- SFS-CNet† reaches 96.2% AP50 on HRSID with 1.86M params and 6.9 GFLOPs, versus 90.8% for YOLOv8s-level baseline row at 10.65M params and 28.4 GFLOPs (SFS-CNet without OGL 88.8 AP-level all-class precision metrics reported in table).
- SFS-CNet† achieves 89.7% mAP on SAR-AIRcraft-1.0, surpassing YOLOv5n by +1.5% and YOLOv8n by +1.3%.
- On SSDD, SFS-CNet† achieves 99.6% AP50, the highest among compared methods, with 8.6 ms inference on HRSID versus 14.1 ms for YOLOv8s.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the OGL strategy | SFS-CNet · HRSID | AP50: 95.7 -> 96.2 | Helped |
| Use the OGL strategy | SFS-CNet · SAR-AIRcraft-1.0 | mAP: 88.7 -> 89.7 | Helped |
| Use the OGL strategy | SFS-CNet · SSDD | AP50: 99.4 -> 99.6 | Helped |
| Set shunt ratio α to 0 instead of α=1/2 (ablation) | SFS-Conv · HRSID test | AP50: 95.71 % -> 94.80 % | Hurt |
| Set shunt ratio α to 1/4 instead of α=1/2 (ablation) | SFS-Conv · HRSID test | AP50: 95.71 % -> 95.73 % | No clear effect |
| Set shunt ratio α to 3/4 instead of α=1/2 (ablation) | SFS-Conv · HRSID test | AP50: 95.71 % -> 95.38 % | Hurt |
| Set shunt ratio α to 1 instead of α=1/2 (ablation) | SFS-Conv · HRSID test | AP50: 95.71 % -> 94.55 % | Hurt |
| Add SPU and FPU instead of ordinary 3×3 convolution (ablation) | SFS-Conv · HRSID test | AP50: 90.39 % -> 95.71 % | Helped |
| Add FPU to SPU-only SFS-Conv (ablation) | SFS-Conv · HRSID test | AP50: 94.66% -> 95.71% | Helped |
| Add SPU to FPU-only SFS-Conv (ablation) | SFS-Conv · HRSID test | AP50: 94.45% -> 95.71% | Helped |
| Add channel selection instead of no feature selection (ablation) | SFS-Conv · HRSID test | AP50: 94.68 % -> 95.71 % | Helped |
| Replace channel selection with spatial selection (ablation) | SFS-Conv · HRSID test | AP50: 95.71 % -> 95.82 % | No clear effect |
Surprising findings
- Replacing SFS-Conv's selection with spatial selection (LSKNet-style) gave the highest AP50 (95.82%) but added parameters and latency, so the parameter-free channel selection was chosen for efficiency.
Limitations
- Gains over the best standard detectors are small on SAR-AIRcraft-1.0 and SSDD, where differences are under about 1 AP.
- The fractional Gabor kernel design and OGL strategy add hyperparameters and design complexity not fully ablated across all datasets.
Experimental setup
Datasets: HRSID SAR-AIRcraft-1.0 SSDD
Benchmarks: HRSID SAR-AIRcraft-1.0 SSDD
Metrics: AP50 AP75 mAP precision recall F1 params FLOPs inference time
Baselines: Faster R-CNN Mask R-CNN Cascade R-CNN HTC RetinaNet FCOS CenterNet YOLOv5n YOLOv5s YOLOv8n YOLOv8s YOLO-FA
Builds on: YOLOv8 Gabor transform Res2Net-style hierarchical residual connections LSKNet DGNet
Closest papers by meaning
- SARDet-100K: Towards Open-Source Benchmark and ToolKit for Large-Scale SAR Object Detection (NeurIPS 2024 · similarity 0.87)
- Dual Information Purification for Lightweight SAR Object Detection (AAAI 2025 · similarity 0.83)
- Statistically Principled Deep Learning for SAR Image Segmentation (AAAI 2024 · similarity 0.82)
- FUSAR-GPT: A Spatiotemporal Feature-Embedded and Two-Stage Decoupled Visual Language Model for SAR Imagery (CVPR 2026 · similarity 0.81)
- HS-FPN: High Frequency and Spatial Perception FPN for Tiny Object Detection (AAAI 2025 · similarity 0.80)
- RSAR: Restricted State Angle Resolver and Rotated SAR Benchmark (CVPR 2025 · 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.