Dataset Statistics Report¶

This report is generated from local dataset files with data_tools/stats/build_dataset_report.py. It gives the project one shared source of truth for image counts, annotation counts, class balance, and representative examples.

What This Report Checks¶

The script verifies four things before training code depends on the datasets:

Check	Input	Output	Normal result
File discovery	Local `datasets/raw/` folders	Image and annotation counts	Counts match the expected public dataset scale.
Label parsing	ELPV CSV and VOC XML files	Class, probability, and box tables	Class names are readable and no parser errors occur.
Long-tail shape	Per-class object counts	Bar charts	Rare classes remain visible instead of being hidden by averages.
Visual sanity	Real image files	Small example grids	Images open correctly and labels match the visible defect type.

Dataset Overview¶

Dataset	Local input	Label format	Images counted	Label units counted
ELPV Dataset	`datasets/raw/elpv-dataset`	CSV defect probability	2624	2624 probability labels
PV-Multi-Defect	`datasets/raw/pv_multi_defect`	Pascal VOC XML boxes	1106	3981 boxes
PVEL-AD	`datasets/raw/pvel_ad/extracted`	Pascal VOC XML boxes	36543	41958 boxes on 23650 annotated images

From ELPV Dataset¶

ELPV is a cell-level electroluminescence image dataset. Each image has one defect-probability label instead of object boxes. The label answers a classification question: how likely is this cell to contain a visible defect?

ELPV generated sample grid

ELPV source overview

ELPV Probability Counts¶

Defect probability	Images
`0.0`	1508
`0.3333333333333333`	295
`0.6666666666666666`	106
`1.0`	715

ELPV Module-Type Counts¶

Module type	Images
`mono`	1074
`poly`	1550

From PV-Multi-Defect¶

PV-Multi-Defect uses object boxes on panel images. It is useful for checking whether a detector can localize visible surface defects, not only classify a full image as defective.

PV-Multi-Defect generated sample grid

Source Example Images From PV-Multi-Defect¶

Broken area source example

Bright spot source example

Border source example

Scratch source example

Non-electricity source example

PV-Multi-Defect Box Counts By Class¶

PV-Multi-Defect class distribution

Class	Boxes
`black_border`	256
`broken`	98
`hot_spot`	2079
`no_electricity`	181
`scratch`	1367

From PVEL-AD¶

PVEL-AD is the main long-tail object detection dataset in this workspace. The detector input is a near-infrared EL image. The output is a set of bounding boxes with one of 12 defect classes.

PVEL-AD generated sample grid

PVEL-AD class distribution

Class	Boxes
`black_core`	4905
`corner`	21
`crack`	4057
`finger`	25596
`fragment`	12
`horizontal_dislocation`	2380
`printing_error`	80
`scratch`	8
`short_circuit`	1707
`star_crack`	218
`thick_line`	2566
`vertical_dislocation`	408

PVEL-AD Split Counts¶

Split	Images	Annotated images	Boxes
trainval	4500	4500	7842
test	19150	19150	34116

PVEL-AD also contains anomaly-free or auxiliary images without VOC boxes. The full local image count is 36543; the table above counts the images that have released VOC annotations.

How To Read The Result¶

The counts are normal when the ELPV image total is close to 2,624, PVEL-AD image total is 36,543, and the object-detection datasets show strong class imbalance. That imbalance is not a data error; it is the reason the training plan needs class-aware sampling, careful recall tracking for rare classes, and separate latency checks after model export.