Research map / Detection and segmentation / CNN features
Acoustic-to-Hyper-Spectral: Hyper-Spectral Image Construction from Frequency Spectrums Through Simulated Annealing (Student Abstract)
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.
Builds hyper-spectral images from DAS frequency spectra by using simulated annealing to train learnable filters that select and bin frequency bands, guided by object detector mAP.
Problem
Constructing informative hyper-spectral images from distributed acoustic sensing (DAS) frequency spectra for object detection, where mapping every frequency to a channel yields oversized images and noisy frequencies must be dropped, while current methods rely on expert knowledge or trial-and-error b
Approach
Learnable filters that split, merge, and drop frequency bands are optimized with simulated annealing, using the downstream detector's COCO mAP as the acceptance criterion, with detector weights transferred across each filter change.
Key results
- Simulated annealing reached 0.793 mAP versus 0.571 for hand-crafted and 0.610 for uniform filters.
- Hill climbing achieved 0.789 and 0.768 mAP across two seeds, showing seed sensitivity.
- Simulated annealing outperformed uniform filtering by 18.3% mAP and hand-crafted by 22.2% mAP per the paper.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace expert-designed filters with simulated-annealing-learned filters | YOLOv7-tiny on the commercial DAS dataset · Commercial DAS dataset | mAP: by a margin of 22.2% | Helped |
| Replace uniform filters with simulated-annealing-learned filters | YOLOv7-tiny on the commercial DAS dataset · Commercial DAS dataset | mAP: by a margin of 18.3% | Helped |
| Use simulated annealing instead of hill climbing for filter search (ablation) | YOLOv7-tiny on the commercial DAS dataset · Commercial DAS dataset | mAP: 0.4% - 2.5% | Helped |
| Change the hill-climb random seed from 132 to 768 (ablation) | YOLOv7-tiny filter search · Commercial DAS dataset | mAP: varies by 2.1% | Hurt |
Surprising findings
- Hill climbing's performance varied by 2.1% mAP across random seeds, indicating high sensitivity to randomness.
Limitations
- Results depend on random seed; search space is restricted because dropped bands cannot be reintroduced.
- Limited action set (split, drop, merge) excludes many frequency combinations.
- Evaluated only on a single proprietary DAS dataset.
Experimental setup
Datasets: Commercial DAS installation dataset (3442 images)
Benchmarks: Proprietary DAS hyper-spectral image test split
Metrics: mAP recall
Baselines: Hand-crafted frequency bands Uniform filters Hill climb
Builds on: YOLOv7
Closest papers by meaning
- Hyper-Spectral Image Generation from Frequency Spectrums (AAAI 2025 · similarity 0.90)
- FRN: Fractal-Based Recursive Spectral Reconstruction Network (NeurIPS 2025 · similarity 0.84)
- Any-Resolution AI-Generated Image Detection by Spectral Learning (CVPR 2025 · similarity 0.84)
- SPECAT: SPatial-spEctral Cumulative-Attention Transformer for High-Resolution Hyperspectral Image Reconstruction (CVPR 2024 · similarity 0.82)
- Hyperspectral Image Reconstruction via Combinatorial Embedding of Cross-Channel Spatio-Spectral Clues (AAAI 2024 · similarity 0.82)
- Balancing Frequencies and Pixels in Flow Matching (NeurIPS 2026 · similarity 0.82)
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.