正規化

Normalization is a fundamental technique in data preprocessing that involves rescaling numeric attributes to a standard range. In the context of machine learning (ML), datasets often contain features with varying scales—such as age ranges (0–100) versus income levels (0–100,000). If left untreated, these disparities can cause the optimization algorithm to become biased toward larger values, leading to slower convergence and suboptimal performance. By normalizing data, engineers ensure that every feature contributes proportionately to the final result, allowing neural networks to learn more efficiently.

一般的な正規化テクニック

There are several standard methods for transforming data, each suited for different distributions and algorithm requirements.

• Min-Max Scaling: This is the most intuitive form of normalization. It rescales the data to a fixed range, usually [0, 1]. This transformation is performed by subtracting the minimum value and dividing by the range (maximum minus minimum). It is widely used in image processing where pixel intensities are known to be bounded between 0 and 255.
• Z-Score Standardization: While often used interchangeably with normalization, standardization specifically transforms data to have a mean of 0 and a standard deviation of 1. This is particularly useful when the data follows a Gaussian distribution and is essential for algorithms like Support Vector Machines (SVM) that assume normally distributed data.
• Log Scaling: For data containing extreme outliers or following a power law, applying a logarithmic transformation can compress the range of values. This makes the distribution more manageable for the inference engine to interpret effectively without being skewed by massive value spikes.

実際のアプリケーション

正規化は、様々な産業における高性能AIシステムのパイプラインにおける標準的なステップである。

1. Computer Vision (CV): In tasks such as object detection and image classification, digital images are composed of pixel values ranging from 0 to 255. Feeding these large integers directly into a network can slow down gradient descent. A standard preprocessing step involves dividing pixel values by 255.0 to normalize them to the [0, 1] range. This practice ensures consistent inputs for advanced models like YOLO26, improving training stability on the Ultralytics Platform.
2. 医療画像解析：医療用スキャン（医療分野におけるAIで使用されるものなど）は、 異なる装置で撮影され、輝度スケールが異なる場合が多い。 正規化処理により、MRIやCTスキャンのピクセル輝度が 異なる患者や装置間で比較可能となる。この一貫性は 腫瘍の正確な検出に不可欠であり、 モデルが輝度の変動ではなく構造的異常を認識できるようにする。

関連概念の区別

It is important to differentiate normalization from similar preprocessing and architectural terms found in deep learning.

• vs. Batch Normalization: Data normalization is a preprocessing step applied to the raw input dataset before it enters the network. Conversely, Batch Normalization operates internally between layers throughout the network during model training. It normalizes the output of a previous activation layer to stabilize the learning process.
• vs. Image Augmentation: While normalization changes the scale of the pixel values, augmentation changes the content or geometry of the image (e.g., flipping, rotating, or changing colors) to increase dataset diversity. Tools like Albumentations are used for augmentation, whereas normalization is a mathematical scaling operation.

実施例

In computer vision, normalization is often the first step in the pipeline. The following Python example demonstrates how to manually normalize image data using the NumPy library, a process that happens automatically within the Ultralytics YOLO26 data loader during training.

import numpy as np

# Simulate a 2x2 pixel image with values ranging from 0 to 255
raw_image = np.array([[0, 255], [127, 64]], dtype=np.float32)

# Apply Min-Max normalization to scale values to [0, 1]
# This standardizes the input for the neural network
normalized_image = raw_image / 255.0

print(f"Original Range: {raw_image.min()} - {raw_image.max()}")
print(f"Normalized Range: {normalized_image.min()} - {normalized_image.max()}")