Image Format

An image is more than a grid of colored pixels: it is also a file with rules for storing, compressing, and interpreting those pixels. An image format is that set of rules—the structure that lets software save an image to disk and later reconstruct it correctly.

What a format controls
Formats define several practical details:

• Pixel encoding: whether colors are stored as RGB, grayscale, or with an alpha channel for transparency.
• Compression: how file size is reduced. Compression can discard detail or preserve every pixel value.
• Metadata: extra information such as camera orientation, color profile, capture time, or resolution.
• Bit depth: how precisely each channel is represented, such as 8-bit values from 0–255 or 16-bit values used in scientific and medical images.

Common choices in vision work
JPEG uses lossy compression, making it compact and useful for ordinary photographs and large web-scale training sets. Its compression can create blocky artifacts or blur fine edges, which matters when detecting tiny defects or reading text. PNG uses lossless compression and supports transparency, so it is well suited to segmentation masks, screenshots, and annotation images where each pixel label must remain exact. TIFF supports high bit depths and lossless storage, making it common in microscopy, medical imaging, and archival workflows. WebP provides efficient lossy or lossless storage, while BMP is simple but usually too large for practical datasets.

Why it matters to models
A model does not directly “understand” JPEG or PNG; a library such as OpenCV or Pillow decodes the file into an array first. Errors arise when that decoding is misunderstood: OpenCV’s cv2.imread() commonly returns BGR channel order rather than RGB, and saving a segmentation mask as JPEG can corrupt its class boundaries. Choosing the right image format preserves the visual evidence—and labels—that a vision system needs to make reliable decisions.