PrepImage

A Unified Framework for Computer Vision Dataset Preparation
Deduplicate, clean, crop, resize, annotate, and augment image datasets — all in one desktop app.

---

Preparing an image dataset for a computer vision or AI project usually means stitching together a pile of one-off scripts — one for finding duplicates, another for resizing, another still for labeling and augmenting. PrepImage replaces all of that with a single, modern PySide6 (Qt) desktop application: open a folder of raw images, and clean, standardize, label, and expand it into a training-ready dataset without leaving the app.

Table of Contents

• Why PrepImage
• Modules
  • 1. Duplicate, Blur & Noise Detection
  • 2. Image Preprocessing
  • 3. Image Annotation
  • 4. Image Augmentation
• Installation
• Usage
• Requirements
• Project Layout
• License
• Contributors

Why PrepImage

• One app, four workflows — deduplication, preprocessing, annotation, and augmentation share one consistent, modern dark-themed interface instead of four different scripts and conventions.
• Built for real datasets — background worker threads keep the UI responsive on folders with thousands of images; every long-running step shows live progress.
• Training-format aware — exports YOLO .txt, Pascal VOC .xml, and COCO-style .json annotations, and augments images with their labels so bounding boxes and polygons stay correct after a flip, rotation, or crop.
• Nothing leaves your machine — PrepImage is a local desktop tool; your images and labels never get uploaded anywhere.

Modules

1. Duplicate, Blur & Noise Detection

Scans a folder of images and flags three kinds of problems that quietly hurt model training if they slip into a dataset:

• Duplicate & near-duplicate images — found via perceptual average-hashing (imagehash), then grouped so you can keep one representative from each group instead of every copy.
• Blurry images — detected via a Laplacian-variance sharpness score.
• Noisy images — detected via a median-filter residual noise estimate.

Enable any combination of the three Detection Options and run one scan; each category gets its own results tab with its own thumbnail grid. From there you can Forward Unique → Preprocess (send the clean set straight into the next module), Download Unique Images, or Download Duplicates for manual review.

Directories

Field	Purpose
Input Directory	The folder of images to scan. Required.
(no separate output field)	Results stay in memory until you explicitly export them — each export action (Forward, Download Unique, Download Duplicates) prompts for its own destination folder when clicked.

2. Image Preprocessing (Crop / Resize / Rename)

Batch-normalizes a folder of images into a consistent size and naming scheme — the shape most training pipelines expect:

• Crops every image to a square, trimming evenly from the longer side (never stretching or distorting the image).
• Resizes the square result to a fixed target size in pixels.
• Renames output files sequentially with a custom prefix.
• Zoomable thumbnail gallery to review and select which images to process before running.

Directories

Field	Purpose
Input Directory	Folder of source images to process. Required.
Output Directory	Folder where cropped/resized/renamed images are written. Required.

3. Image Annotation

A full labeling workspace for building object detection and segmentation datasets:

• Bounding boxes for object detection and polygons for pixel-level segmentation, with an optional rasterized mask PNG export.
• Per-image class labels, editable at any time (a "⋮" menu on every class and every annotation row covers editing, recoloring, renaming, changing class, and deleting).
• Undo/redo — whole-shape while browsing, per-vertex while a polygon is still being drawn.
• Export to YOLO .txt, Pascal VOC .xml, and/or COCO-style .json — pick any combination — with class names round-tripped through a per-folder classes.txt so relabeling later doesn't lose your class names.
• Zoom/pan canvas, keyboard shortcuts, and an in-app shortcuts reference.

Directories

Field	Purpose
Input Image Folder	Folder of images to annotate. Required.
Custom Save Directory	Where annotation files (and masks, if enabled) are saved. Optional — defaults to the input image folder if left blank.

4. Image Augmentation

Expands a dataset by generating pixel-level and geometric variations of each image, in two modes:

• Image Augmentation — augments a folder of images with no annotations to carry along.
• Annotated Image Augmentation — augments images and their existing YOLO .txt annotations together. Pixel/non-geometric augmentations (brightness, contrast, sharpen, Gaussian blur, saturation, Gaussian noise, grayscale) leave box coordinates untouched; geometric ones (horizontal/vertical flip, 90°/180°/270° rotation) transform the coordinates to match, so every saved label still lines up with its image.

Target Images controls how many augmented images to generate in total; Max Aug / Img caps how many augmented variants can come from any single source image. Augmented files follow the naming convention originalname_SHORTCODE1_SHORTCODE2_0001.ext, so it's obvious at a glance which combination of augmentations produced a given output.

Directories

Field	Purpose
Input Image Folder	Folder of source images (in Annotated mode, this folder should also contain the matching .txt label for each image). Required.
Output Image Folder	Folder where augmented images (and, in Annotated mode, their updated .txt labels) are written. Required.

Installation

From PyPI

pip install prepimage

From source (editable / dev mode)

git clone https://github.com/harsh-iasri/PrepImage.git
cd prepimage/prepimage_pkg
pip install -e .

Usage

prepimage

or, without installing the console script:

python -m prepimage

Requirements

• Python >= 3.9
• PySide6 — Qt GUI framework
• Pillow — image I/O and processing
• ImageHash — perceptual hashing for duplicate detection
• NumPy — array operations for preprocessing, blur/noise detection, and augmentation

All of the above are installed automatically as dependencies.

Project Layout

prepimage_pkg/
├── pyproject.toml
├── README.md
├── LICENSE
└── src/
    └── prepimage/
        ├── __init__.py
        ├── __main__.py            # entry point (`prepimage` / `python -m prepimage`)
        ├── app.py                 # QMainWindow + navigation (QStackedWidget)
        ├── theme.py               # colors, fonts, global stylesheet
        ├── widgets.py             # ToolCard (painted card, hover zoom)
        ├── assets.py              # shared logo loading helper
        ├── duplicates.py          # duplicate/blur/noise detection core logic
        ├── preprocess.py          # crop/resize/rename core logic
        ├── augment.py             # augmentation engine + box/polygon transforms
        └── screens/
            ├── __init__.py
            ├── home.py             # landing page (hero + 4 tool cards + About dialog)
            ├── duplicates.py       # Duplicate, Blur & Noise Detection screen
            ├── preprocess.py       # Preprocessing screen
            ├── annotate.py         # Image Annotation screen (boxes + polygons)
            ├── augment.py          # Image Augmentation screen
            ├── working.py          # shared "under construction" placeholder
            └── images/             # screenshots + logo bundled with the package

License

Released under the MIT License.

Contributors

Name	Affiliation
Harsh Sachan	ICAR-Indian Agricultural Statistics Research Institute (IASRI), New Delhi
Shalini Kumari	ICAR-Indian Agricultural Statistics Research Institute (IASRI), New Delhi
Md Ashraful Haque	ICAR-Indian Agricultural Statistics Research Institute (IASRI), New Delhi
Sudeep Marwaha	ICAR-Central Institute of Agricultural Engineering (CIAE), Bhopal
Chandan Kumar Deb	ICAR-Indian Agricultural Statistics Research Institute (IASRI), New Delhi

Contributions are welcome — feel free to open an issue or pull request on GitHub.