Skip to content

Classification & Detection Dataset

The Classification Tool is designed to build the training datasets required by deep learning classification models. It supports real-time data augmentation before export, multi-format transcoding, and flexible sample map partitioning, allowing you to generate a directory structure that meets the requirements of mainstream frameworks with a single click.

class.png

1. Core Workflow

The classification tool performs its operations in the following logical order:

  1. Read: Scans the original images in the input directory, where images are already grouped by folder class.

  2. Augment: Applies mirroring, rotation, or color transformations to the samples of each class according to the configuration.

  3. Transcode: Converts all generated samples uniformly into the specified image format.

  4. Distribute: Splits the data into the train (training set) and val (validation set) folders according to the configured ratio.


2. Data Augmentation Configuration

Before generation, you can check the following augmentation operators. Based on the original images, the system expands them by performing transformations up to the "Image Count" you set:

Augmentation Operator Physical Logic Industrial Application Recommendation
Horizontal Mirroring Left-right flip (symmetric about the Y axis). Suitable for distinguishing left/right variants of a workpiece.
Vertical Mirroring Top-bottom flip (symmetric about the X axis). Suitable for conveyor materials with no fixed orientation.
Full-Turn Rotation 360° rotation at a random angle. Essential for round workpieces; solves the rotation invariance problem for dot-marking templates.
Hue / Contrast Random brightness and color shift. Simulates lighting fluctuations and camera exposure differences on the production floor.

3. Generation Parameters

Image Output Format

You can choose the target format according to your storage requirements or the requirements of the downstream training framework:

  • BMP / PNG: Lossless formats, recommended for high-precision industrial defect classification.

  • JPG: Compressed format, suitable for large-scale datasets and saves disk space.

  • Original format: No re-encoding; the original suffix of the input files is kept as is.

Training / Validation Set Ratio

  • How it works: The ratio value represents the percentage of data assigned to the training set.
  • Split example: If set to 80, 80% of the data goes into the train folder and 20% goes into the val folder.
  • Special setting: If set to 100, the system performs no splitting and outputs everything as the training set.

4. Steps

  1. Configure the paths:
    • Input directory: Select the root path containing the original class folders (such as OK/, NG/).
    • Output directory: Specify where the generated dataset will be stored (an empty new folder is recommended).
  2. Set the image count: Enter the total number of images you ultimately want generated for each class.
  3. Run the generation: After confirming the format and the ratio, click [Start Generation].

classdata.png


5. Preview of the Output Directory Structure

After the generation is complete, your output directory will have the following structure:

classcomplete.png

dataset_20260415_114920/

├── train/              # Training set path (Train)

│   ├── OK/       # Augmented normal samples

│   └── NG/       # Augmented defect samples

└── val/                # Validation set path (Val)

    ├── OK/       # Automatically split validation samples

    └── NG/       # Automatically split validation samples

├── Label         # Contains class information

├── train_list    # Image list of the training set

├── val_list      # Image list of the validation set

Sharing from Development Practice

In industrial classification tasks, sample balance is critical. If you only have 20 original NG samples but 200 OK samples, augment the NG class to roughly 200 images, which effectively prevents the model from developing a class bias.