How to do image segmentation without machine learning?

Image segmentation without machine learning can be achieved using traditional computer vision techniques that rely on pixel properties, mathematical operations, or manual feature engineering. These methods focus on analyzing color, intensity, texture, or edges to partition images into meaningful regions. While less automated than ML-based approaches, they are computationally efficient and transparent, making them useful for specific applications with predictable patterns or limited computational resources.

One common approach is thresholding, which separates pixels based on intensity values. For example, Otsu’s method automatically calculates an optimal threshold to distinguish foreground from background by analyzing the image histogram. Another technique is edge detection using operators like Sobel, Canny, or Prewitt to identify abrupt intensity changes. Once edges are detected, morphological operations (e.g., dilation, erosion) can close gaps and form contiguous boundaries. Region-based methods like region growing start with seed pixels and expand regions by merging adjacent pixels that meet similarity criteria (e.g., intensity within a tolerance). Similarly, the watershed algorithm treats pixel intensity as elevation and simulates flooding from markers to segment overlapping objects. These methods work best in scenarios with clear contrast or distinct edges, such as separating cells in microscopy images.

For color-based segmentation, techniques like color space thresholding (e.g., isolating green regions in HSV space) or clustering (e.g., K-means on RGB values) can partition images without requiring training data. Graph-based methods like graph cuts define energy functions to separate regions based on user-defined foreground/background seeds. However, these approaches often require manual parameter tuning (e.g., threshold values, seed placement) and struggle with complex textures or noisy data. Tools like OpenCV provide built-in functions for many of these techniques, such as cv2.threshold for basic segmentation or cv2.watershed for marker-based workflows. While less flexible than ML models, these methods remain practical for constrained problems where simplicity and speed are prioritized over adaptability.