Research map / Detection and segmentation / Open-set and few-shot detection
Towards Automatic Power Battery Detection: New Challenge Benchmark Dataset and Baseline
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.
Introduces the power battery detection (PBD) task, an X-ray PBD dataset of 1,500 images, and MDCNet, a segmentation-based multi-task network with a prompt filter, line and counting predictors, and distance-adaptive label generation.
Problem
Automatically localizing dense cathode and anode plate endpoints in X-ray images of power batteries to evaluate battery quality (number and overhang), which is currently done by human inspection and is slow and error-prone.
Approach
Casts plate endpoint localization as point segmentation, aided by auxiliary line segmentation and counting predictors, a prompt filter module that uses a clean reference image to suppress interference, and a distance-adaptive point-mask label generation strategy.
Key results
- MDCNet reaches AN-ACC 0.9633, CN-ACC 0.9908 and PN-ACC 0.9541 on the regular split, versus 0.6881 PN-ACC for the second-best CFINet.
- On the difficult and tough splits MDCNet attains PN-ACC 0.7603 and 0.5115, the best among compared methods.
- Distance-adaptive labels (Ada-0.3) outperform fixed-radius constant labels; PFM with prompt images shows stable results across five prompt choices.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MDCNet instead of CFINet | X-ray PBD detection · X-ray PBD regular | PN-ACC: 0.6881 -> 0.9541 | Helped |
| Use MDCNet instead of CFINet | X-ray PBD detection · X-ray PBD difficult | PN-ACC: 0.5426 -> 0.7603 | Helped |
| Use MDCNet instead of CFINet | X-ray PBD detection · X-ray PBD tough | PN-ACC: 0.3276 -> 0.5115 | Helped |
| Add a prompt filter module (ablation) | point-segmentation baseline · X-ray PBD regular | PN-ACC: 0.6239 -> 0.7706 | Helped |
| Add a counting predictor (ablation) | MDCNet with PFM · X-ray PBD regular | PN-ACC: 0.7706 -> 0.8716 | Helped |
| Add a line predictor (ablation) | MDCNet with PFM and counting predictor · X-ray PBD regular | PN-ACC: 0.8716 -> 0.9541 | Helped |
| Use distance-adaptive point masks with diameter 0.3x adjacent-plate distance instead of fixed-radius masks (ablation) | MDCNet point-segmentation branch · X-ray PBD regular | PN-ACC: Const-1 0.4587; Const-3 0.7523; Const-5 0.7339; Ada-0.1 0.6055; Ada-0.3 0.9541; Ada-0.5 0.8991 | Helped |
| Vary the P-attribute prompt image used by PFM (ablation) | MDCNet · X-ray PBD regular | PN-ACC: 0.9541 | No clear effect |
Surprising findings
- The prompt filter module gives results that are essentially identical across five different prompt images, suggesting insensitivity to the specific prompt choice.
Limitations
- Performance on difficult and tough splits remains poor; dense, overlapping plates still cause adhesions.
- Annotation relies on experienced technicians, so data is scarce.
- Only 2D X-ray data; 3D CT extension is future work.
Experimental setup
Datasets: X-ray PBD
Benchmarks: X-ray PBD regular split X-ray PBD difficult split X-ray PBD tough split
Metrics: AN-MAE CN-MAE AN-ACC CN-ACC PN-ACC AL-MAE CL-MAE OH-MAE
Baselines: Harris corner detector Shi-Tomasi Sub-pixel corner detection BL (Bayesian loss counting) CUT IOCFormer RetinaNet Faster R-CNN YOLOv5 C3Det CFINet
Builds on: FPN ResNet-50 Bayesian loss
Closest papers by meaning
- BatteryML: An Open-source Platform for Machine Learning on Battery Degradation (ICLR 2024 · similarity 0.82)
- NOCE-Net: Representation Learning for Battery Operational Context via Nested Sequence Modelling (NeurIPS 2026 · similarity 0.77)
- PowerGraph: A power grid benchmark dataset for graph neural networks (NeurIPS 2024 · similarity 0.73)
- PowerPM: Foundation Model for Power Systems (NeurIPS 2024 · similarity 0.73)
- Energy Efficient Streaming Time Series Classification with Attentive Power Iteration (AAAI 2024 · similarity 0.71)
- CableInspect-AD: An Expert-Annotated Anomaly Detection Dataset (NeurIPS 2024 · similarity 0.71)
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.