Research map / Architectures and efficiency / Attention and transformers
IRPruneDet: Efficient Infrared Small Target Detection via Wavelet Structure-Regularized Soft Channel Pruning
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 IRPruneDet, a soft channel pruning method for infrared small target detection that uses wavelet-domain (Haar) l1 regularization for channel importance and soft channel reconstruction to avoid premature pruning, cutting U-Net18 parameters and FLOPs by about 64% and 51% while improving IoU.
Problem
Infrared small target detection (IRSTD) models are large and computationally expensive, but standard channel pruning criteria (magnitude-based) and hard iterative pruning discard channels carrying faint target information under low SNR.
Approach
Represent conv filters in the wavelet domain and apply l1 regularization on Haar wavelet coefficients (approximated as a gradient/differential operator to avoid extra memory) to score channels, combined with soft channel reconstruction that stochastically re-activates previously pruned channels by interpolating with the best-performing parameters under a cosine schedule.
Key results
- With U-Net18 baseline, IRPruneDet reduces parameters by 64.13% and FLOPs by 51.19%, raising IoU from 73.31% to 75.12% and nIoU from 70.92% to 74.30% on NUAA-SIRST.
- On IRSTD-1k, IRPruneDet reports 64.54% IoU and 91.74% Pd with 0.938G FLOPs and 0.1802M parameters, versus FC3-Net at 64.98% IoU and 10.57G FLOPs / 6.896M parameters.
- Ablation: WCP+SCR (75.12 IoU) outperforms l1-norm+SCR (73.51) and WCP alone (74.25), and beats SFP, FPGM and ASFP pruning baselines at 50% sparsity.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use WCP instead of l1-norm channel importance pruning (ablation) | U-Net pruning on NUAA-SIRST · NUAA-SIRST | IoU: 67.18 -> 74.25 | Helped |
| Add SCR to l1-norm pruning (ablation) | U-Net pruning on NUAA-SIRST · NUAA-SIRST | IoU: 67.18 -> 73.51 | Helped |
| Add SCR to WCP pruning (ablation) | U-Net pruning on NUAA-SIRST · NUAA-SIRST | IoU: 74.25 -> 75.12 | Helped |
| Replace IRPruneDet pruning with general pruning methods (ablation) | U-Net pruning on NUAA-SIRST · NUAA-SIRST | IoU: 75.12 vs 67.18; 70.71; 71.54; 70.82; 71.86 | Hurt |
| Use IRPruneDet instead of FC3-Net | IRSTD models · NUAA-SIRST and IRSTD-1k | IoU: 74.75 -> 75.12; 64.98 -> 64.54 | Mixed |
| Use IRPruneDet instead of FC3-Net | IRSTD models · NUAA-SIRST | FLOPs (10^9): 10.57 -> 0.9380 | Helped |
Surprising findings
- Conventional l1-norm magnitude pruning reduces IoU (73.31 to 67.18) on the U-Net baseline, suggesting magnitude criteria discard low-magnitude but target-relevant channels.
Limitations
- Evaluated only on two single-frame IRSTD datasets with a U-Net18 baseline.
- Relies on hand-tuned hyperparameters such as lambda, delta-beta SCR and beta0 (default values reported without broad sensitivity analysis).
Experimental setup
Datasets: NUAA-SIRST IRSTD-1k
Benchmarks: NUAA-SIRST IRSTD-1k
Metrics: IoU nIoU Pd Fa FLOPs parameters
Baselines: MDvsFA ACM ALCNet FC3-Net Top-Hat Max-Median WSLCM TLLCM IPI NRAM RIPT PSTNN
Builds on: soft filter pruning (SFP) U-Net ACM Haar wavelet tight frame
Closest papers by meaning
- Rethinking Evaluation of Infrared Small Target Detection (NeurIPS 2025 · similarity 0.88)
- Rethinking Infrared Small Target Detection: A Foundation Driven Efficient Paradigm (NeurIPS 2026 · similarity 0.88)
- Semi-supervised Infrared Small Target Detection with Thermodynamic-Inspired Uneven Perturbation and Confidence Adaptation (AAAI 2025 · similarity 0.87)
- SAIST: Segment Any Infrared Small Target Model Guided by Contrastive Language-Image Pretraining (CVPR 2025 · similarity 0.86)
- IRMamba: Pixel Difference Mamba with Layer Restoration for Infrared Small Target Detection (AAAI 2025 · similarity 0.86)
- Infrared Small Target Detection with Scale and Location Sensitivity (CVPR 2024 · similarity 0.86)
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.