🟧 Image Normalization

Function Description

This operator is used to readjust the brightness information of each pixel in the input image, unifying pixel values of different scales or distributions to a standard baseline, eliminating numerical differences caused by different brightness, contrast, or bit depth between images.

Usage Scenarios

  • Deep learning models may require input image pixel values to be within a standard range (such as [0, 1] or [-1, 1]), or conform to standard normal distribution (mean 0, variance 1), which can be preprocessed through image normalization.

  • When images are captured under different lighting conditions, it may cause inconsistent image brightness. Normalization can reduce the impact of lighting changes and improve algorithm robustness.

Input Output

Input

Image: Input images or image lists to be processed, can be grayscale or color images.

Output

Normalized Image: Images or image lists after normalization processing.

Parameter Description

Normalization Method

Parameter Description

Select the algorithm for normalizing image pixel values.

Parameter Adjustment

  • Min-Max Normalization: This method can maintain the distribution structure of original data, performing linear stretching of pixel values and mapping them to a specific target range. Suitable for general scenarios that need to preserve relative relationships between pixel values.

  • Mean-Variance Normalization: Convert pixel values to a standardized distribution, usually with mean 0 and standard deviation 1, then adjust data scale by multiplying by the maximum value of the normalization range. Commonly used in deep learning, as many models assume input data follows standard normal distribution.

  • Max Absolute Value Normalization: Find the largest absolute value pixel in the image, divide each pixel by this value, and scale to the specified range. This method does not change the "center" of the image, i.e., the relative distribution of positive and negative values.

  • Logarithmic Normalization: When brightness differences in the image are very large (e.g., small amount of bright pixels and large amount of dark pixels), logarithmic normalization takes the logarithm of each pixel value, which can reduce differences in bright areas and enhance details in dark areas, making the overall image look more balanced.

Normalization Range Maximum Value

Parameter Description

Set the maximum value or scaling factor of the numerical range after normalization.

Parameter Adjustment

  • When using min-max normalization, this parameter is used to set the upper limit of output range. For example, set to 255, output range is [0, 255], suitable for direct display as 8-bit images; set to 1, output range is [0, 1], commonly used for deep learning.

  • For other normalization methods, this parameter serves as a multiplicative scaling factor applied after normalization calculation.

Parameter Range

[1,10000000], Default value: 1

Enable Clipping

Parameter Description

Control whether to force limit normalization results within a specified minimum and maximum value range.

Parameter Adjustment

  • Off (Default): No clipping performed.

  • On: Enable clipping function. When normalization results may produce extreme values beyond expected range, enabling this option can effectively suppress these outliers, ensuring output values are between minimum and maximum values.

Enable Node

Parameter Description

Control whether this operator executes operations.

Parameter Adjustment

  • On (Default): Normally run the operator function.

  • Off: The operator does not perform any operations and directly outputs the input data.