| --- |
| task_categories: |
| - object-detection |
| tags: |
| - pcb |
| - electronics |
| - yolo |
| - component-detection |
| pretty_name: PCB Components Labelled |
| --- |
| |
| # PCB Components Labelled |
|
|
| YOLO-format object detection datasets and trained weights for detecting electronic |
| components on PCB (printed circuit board) images. Two dataset/model pairs are |
| included, covering a coarse 5-class label set and a fine-grained 17-class label set. |
|
|
| ## Contents |
|
|
| ``` |
| pcb_final_v4/ 17-class dataset (train/valid, YOLO format) |
| pcb_5class_tiled/ 5-class dataset, tiled images (train/valid, YOLO format) |
| models/17_class_pcb_v5_safe/ best.pt trained on pcb_final_v4 |
| models/5_class_pcb_5class_v1/ best.pt trained on pcb_5class_tiled |
| ``` |
|
|
| ## Dataset: `pcb_final_v4` (17 classes) |
|
|
| `data.yaml`: |
| ``` |
| nc: 17 |
| names: ['IC', 'capacitor', 'connector', 'crystal_oscillator', 'diode', 'ferrite_bead', |
| 'fuse', 'inductor', 'led', 'resistor', 'switch', 'transistor', 'button', 'em', |
| 'pins', 'pads', 'test_point'] |
| ``` |
|
|
| ## Dataset: `pcb_5class_tiled` (5 classes) |
|
|
| `data.yaml`: |
| ``` |
| names: |
| 0: resistor |
| 1: capacitor |
| 2: ic |
| 3: led |
| 4: diode |
| ``` |
| Images are tiled crops (higher resolution training strategy), which is why this |
| dataset has far more files than `pcb_final_v4` despite covering fewer classes. |
|
|
| **Note: this dataset is only partially uploaded (~40%).** The full local dataset |
| has 27,274 train images / 13,637 train labels / 6,934 valid images / 3,467 valid |
| labels (~51k files, ~22.5 GB). `train/images` and `train/labels` both exceed the |
| Hugging Face Hub's 10,000-files-per-directory git limit, which blocked further |
| uploads once each hit its cap (currently ~9,989 and ~9,999 files committed |
| respectively). `valid/images` and `valid/labels` are under the limit but were not |
| yet fully uploaded either. The `5_class_pcb_5class_v1` model below was trained on |
| the complete local dataset, not this partial upload — the model's reported metrics |
| still apply in full. |
|
|
| ## Models |
|
|
| Both are Ultralytics YOLO checkpoints (`best.pt`), trained for 150/150 epochs. |
|
|
| | Model | Dataset | Classes | mAP50 | mAP50-95 | Precision | Recall | |
| |---|---|---|---|---|---|---| |
| | `models/17_class_pcb_v5_safe/best.pt` | pcb_final_v4 | 17 | 0.709 | 0.511 | 0.736 | 0.683 | |
| | `models/5_class_pcb_5class_v1/best.pt` | pcb_5class_tiled | 5 | 0.857 | 0.565 | 0.854 | 0.780 | |
|
|
| `args.yaml` and `results.csv` from the corresponding training run are included |
| alongside each `best.pt` for reproducibility. |
|
|
| ## Usage |
|
|
| ```python |
| from ultralytics import YOLO |
| |
| model = YOLO("models/17_class_pcb_v5_safe/best.pt") |
| results = model.predict("your_pcb_image.jpg") |
| ``` |
|
|
| ## License |
|
|
| Not yet specified by the dataset owner. |
|
|
