Extract Features

Intensity features represent the average brightness of each cell, normalized from 0 (black) to 1 (white). This is the most commonly used feature type for document classification, as it captures the overall layout and presence of content in each region.

• Use for most structured documents, such as forms, invoices, and letters.
• Provides a simple, robust basis for distinguishing between document types with different layouts.
• Fast to compute and hardware-accelerated.

GradientAngle features capture the dominant edge direction within each cell, indicating the orientation of lines, text, or graphical elements.

• Useful for distinguishing documents with different line or text orientations.
• Most effective on grayscale or color images, especially photographs or documents with significant graphical content.
• May not add value for simple black-and-white forms.

COG_Angle features measure the direction from the center of each cell to the center of gravity (COG) of dark pixels within the cell. This captures the spatial distribution of content, such as the alignment of text or graphics.

• Combine with 'Intensity' for improved accuracy on forms with graphical elements, stamps, or logos.
• Helps distinguish between document types with similar layouts but different content positioning.

COG_Magnitude features measure the normalized distance from the center of each cell to the center of gravity of dark pixels. This quantifies how far content is offset from the cell center.

• Useful for detecting the presence and spread of content within each cell.
• Can help differentiate between filled and empty regions, or between different types of marks.

Entropy features represent the local complexity or "busyness" of each cell, based on edge and gradient information. High entropy indicates more variation, such as handwriting, stamps, or background noise.

• Use to distinguish between blank, simple, and complex regions.
• Helpful for detecting signatures, stamps, or noisy backgrounds.
• May increase feature vector size and processing time.

ColorSpaceChannel features compute the average value of a selected color channel (e.g., Red, Green, Blue, Hue, Saturation, Value) for each cell.

• Use when color information is important for classification, such as distinguishing between highlighted, stamped, or color-coded documents.
• Requires configuration of the 'Color Space' and 'Channel' properties.
• May not be relevant for black-and-white or grayscale documents.

• Channel1: Red – Intensity of the red component.
• Channel2: Green – Intensity of the green component.
• Channel3: Blue – Intensity of the blue component.

Use RGB when you want to work directly with the original color components of the image.

• Channel1: Hue – The dominant color type (measured as an angle on the color wheel).
• Channel2: Saturation – The intensity or purity of the color.
• Channel3: Value – The brightness or intensity of the color.

HSV is useful for isolating color features or separating color from intensity.

• Channel1: L (Luminosity) – Lightness, ranging from black to white.
• Channel2: A – Green to magenta color component.
• Channel3: B – Blue to yellow color component.

LAB is ideal for separating lightness from color, making it useful for thresholding and feature extraction.

• Channel1: Y – Luma (brightness).
• Channel2: Cb – Chroma blue-difference (blue color information).
• Channel3: Cr – Chroma red-difference (red color information).

YCbCr is commonly used in video and image compression, and is useful for separating brightness from color.

• Channel1: Y – Luma (brightness).
• Channel2: U – Chroma blue-difference (color information).
• Channel3: V – Chroma red-difference (color information).

YUV is often used in television and video systems, and is effective for separating intensity from color.

• Channel1: X – A mix of cone response curves (not directly visible).
• Channel2: Y – Luminance (perceived brightness).
• Channel3: Z – A mix of cone response curves (not directly visible).

XYZ is a device-independent color space used as a basis for other color models.

• Channel1: Hue – The dominant color type.
• Channel2: Lightness – The brightness or intensity of the color.
• Channel3: Saturation – The intensity or purity of the color.

HLS is similar to HSV but uses lightness instead of value, and is useful for color-based feature extraction.