Implementation:Datajuicer Data juicer ImageSubplotFilter
| Knowledge Sources | |
|---|---|
| Domains | Image Filtering, Computer Vision, Hough Transform |
| Last Updated | 2026-02-14 16:00 GMT |
Overview
Detects and filters out image samples that contain subplot/grid layouts by analyzing their internal line structure using Hough Line Transform and edge detection.
Description
This filter is valuable for cleaning image datasets of composite/grid images (e.g., comparison charts, multi-panel figures) that may confuse vision models expecting single coherent images.
Algorithm: 1. Convert images to grayscale and apply Gaussian blur for noise reduction 2. Apply Canny edge detection with configurable thresholds 3. Use Hough Line Transform (probabilistic variant) to detect straight lines 4. Classify detected lines as horizontal or vertical based on angle tolerance 5. Calculate a confidence score from multiple weighted components:
- Line count score (20% each for H/V) -- Ratio of detected lines to minimum thresholds
- Regularity score (15% each for H/V) -- Line spacing consistency via coefficient of variation
- Grid structure score (20%) -- Intersection density analysis
- Length consistency score (5% each for H/V) -- Line length uniformity
Class: ImageSubplotFilter Extends the Filter base class with `_batched_op = True` for efficient batch processing. Key methods:
- compute_stats_single() -- Detects subplots in each image and stores confidence scores, horizontal/vertical line counts, and a boolean subplot_detected flag in the sample stats.
- process_single() -- Applies threshold checks combining confidence score, minimum horizontal lines, and minimum vertical lines. Supports "any" (filter if any image has subplots) and "all" (filter only if all images have subplots) strategies.
Usage
Configure in YAML to filter out images containing grid-like subplot layouts. Tune Canny and Hough parameters for different image types.
Code Reference
Source Location
- Repository: Datajuicer_Data_juicer
- File: data_juicer/ops/filter/image_subplot_filter.py
- Lines: 1-392
Signature
@OPERATORS.register_module("image_subplot_filter")
@LOADED_IMAGES.register_module("image_subplot_filter")
class ImageSubplotFilter(Filter):
_batched_op = True
def __init__(
self, min_horizontal_lines: int = 3, min_vertical_lines: int = 3,
min_confidence: float = 0.5, any_or_all: str = "any",
canny_threshold1: int = 70, canny_threshold2: int = 190,
hough_threshold: int = 110, min_line_length: int = 110,
max_line_gap: int = 18, angle_tolerance: float = 4.0,
*args, **kwargs,
): ...
def compute_stats_single(self, sample, context=False): ...
def process_single(self, sample) -> bool: ...
Import
from data_juicer.ops.filter.image_subplot_filter import ImageSubplotFilter
I/O Contract
Inputs
| Name | Type | Required | Description |
|---|---|---|---|
| min_horizontal_lines | int | No | Minimum horizontal lines for subplot detection (default: 3) |
| min_vertical_lines | int | No | Minimum vertical lines for subplot detection (default: 3) |
| min_confidence | float | No | Minimum confidence score threshold (default: 0.5) |
| any_or_all | str | No | Strategy for multi-image samples: "any" or "all" (default: "any") |
| canny_threshold1 | int | No | First Canny edge detection threshold (default: 70) |
| canny_threshold2 | int | No | Second Canny edge detection threshold (default: 190) |
| hough_threshold | int | No | Hough Transform accumulator threshold (default: 110) |
| min_line_length | int | No | Minimum detectable line length in pixels (default: 110) |
| max_line_gap | int | No | Maximum gap between line segments (default: 18) |
| angle_tolerance | float | No | Tolerance in degrees for H/V classification (default: 4.0) |
Outputs
| Name | Type | Description |
|---|---|---|
| keep | bool | True to retain the sample, False to filter it out |
| stats.image_subplot_confidence | List[float] | Confidence scores per image |
| stats.horizontal_peak_count | List[int] | Horizontal line counts per image |
| stats.vertical_peak_count | List[int] | Vertical line counts per image |
| stats.subplot_detected | bool | Whether any image contains detected subplots |
Usage Examples
# In YAML config:
# process:
# - image_subplot_filter:
# min_horizontal_lines: 3
# min_vertical_lines: 3
# min_confidence: 0.5
# any_or_all: 'any'
# canny_threshold1: 70
# canny_threshold2: 190
# Programmatic usage:
from data_juicer.ops.filter.image_subplot_filter import ImageSubplotFilter
filter_op = ImageSubplotFilter(
min_horizontal_lines=3,
min_vertical_lines=3,
min_confidence=0.6,
any_or_all="any",
)
# Compute stats and filter
sample = filter_op.compute_stats_single(sample, context=True)
keep = filter_op.process_single(sample)
print(f"Keep sample: {keep}")
print(f"Confidence: {sample['stats']['image_subplot_confidence']}")