Cloud Vision API is Google Cloud's core image-understanding service, exposing eleven standard features. LABEL_DETECTION returns open-vocabulary object and scene tags. TEXT_DETECTION runs scene OCR for short strings in natural images. DOCUMENT_TEXT_DETECTION targets documents and handwriting, returning a hierarchical block / paragraph / word / symbol structure. FACE_DETECTION returns face landmarks and emotion likelihoods (joy, sorrow, anger, surprise) but never an identity. LANDMARK_DETECTION, LOGO_DETECTION, and OBJECT_LOCALIZATION cover landmarks, brand logos, and bounded objects. IMAGE_PROPERTIES returns dominant colors and crop hints. SAFE_SEARCH_DETECTION scores adult / spoof / medical / violence / racy on a five-step UNLIKELY-to-VERY_LIKELY scale. CROP_HINTS suggests smart crops. WEB_DETECTION runs a reverse-image lookup and returns a best-guess label. Billing is feature-times-image: one image with LABEL_DETECTION + TEXT_DETECTION bills 2 units. The free tier covers the first 1,000 units per month per feature. From 1,001 to 5,000,000 units per month, most features run $1.50 per 1,000 units; OBJECT_LOCALIZATION is $2.25, WEB_DETECTION is $3.50, and CROP_HINTS is $0.60. Above 5M units per month, TEXT_DETECTION, DOCUMENT_TEXT_DETECTION, FACE, LANDMARK, and LOGO drop to $0.60, and LABEL_DETECTION drops to $1.00. There are four endpoint shapes: v1/images:annotate (synchronous, up to 16 images per batch), v1/images:asyncBatchAnnotate (async with output to GCS), v1/files:annotate (synchronous PDF/TIFF, 5 pages max), and v1/files:asyncBatchAnnotate (async PDF/TIFF up to 2,000 pages per document). Image input takes three forms: inline base64 (HTTP payload capped at 10MB), GCS URI (single file capped at 20MB), or HTTP/HTTPS URL (demo-only). The default project quota is 1,800 requests per minute; higher concurrency needs a Cloud Quotas adjustment. Authentication follows the Google Cloud stack: ADC, service account JSON, OAuth bearer, or a limited API key. Seven official client libraries are maintained: C#, Go, Java, Node.js, PHP, Python, Ruby. Two gotchas dominate field experience. First, teams used to per-request pricing under-budget because Vision bills per feature per image — stacking three features triples the unit count. Decide which features you actually need before estimating cost. Second, multi-page PDFs should not be looped through synchronous TEXT_DETECTION one page at a time. Use DOCUMENT_TEXT_DETECTION via files:asyncBatchAnnotate and let Google process the whole document in one async job that writes JSON output to GCS. For 1:N face identification, Vision is the wrong tool — switch to AWS Rekognition or a self-hosted face model.