--- title: "AI Vision" slug: "ai-vision" updated: 2026-07-13T15:45:51Z published: 2026-07-13T15:45:51Z canonical: "azure-cost-management-playbook.turbo360.com/ai-vision" --- > ## Documentation Index > Fetch the complete documentation index at: https://azure-cost-management-playbook.turbo360.com/llms.txt > Use this file to discover all available pages before exploring further. # AI Vision ## What it does Analyses the visual content of images and video. Core capabilities include: - **Image Analysis** — detects objects, scenes, colours, brands, and generates captions. - **OCR (Read API)** — extracts printed and handwritten text from images and PDFs. - **Face** — detects, verifies, and identifies human faces (subject to responsible AI access controls). - **Spatial Analysis** — counts people and tracks movement in video streams (requires on-premises container). ## When you'd use it - You have image or video content and need to extract structured information without training a custom model. - You need to make images searchable (auto-tagging) or need to moderate visual content. - You are processing scanned documents at scale and need to extract text quickly. ## Sample use case A retail chain wants to automatically tag product images uploaded by suppliers so that items are discoverable in their e-commerce catalogue. Azure AI Vision's Image Analysis API tags each image with object labels and dominant colours; the results are stored in Azure AI Search to power filtered product browsing. ## Example reference architecture ![](https://cdn.document360.io/4b481a5c-c963-4b45-80c8-bc19500830f4/Images/Documentation/image(75).png) ## Pricing & How Costs Work Azure AI Vision is priced **per 1,000 transactions** (API calls). Each feature is metered independently, and different features are priced at different rates: | **Feature** | **Pricing unit** | **Notes** | | --- | --- | --- | | **Image Analysis (tags, objects, captions, brands, colours)** | Per 1,000 transactions | One API call = one image = one transaction | | **Image Analysis — dense captioning** | Per 1,000 transactions | Higher rate than standard analysis; generates captions for regions within the image | | **OCR (Read API — images)** | Per 1,000 transactions | One transaction = one image submitted | | **OCR (Read API — PDF pages)** | Per 1,000 pages | PDF pages are billed per page, not per document | | **Face — detection** | Per 1,000 transactions | Detecting faces in an image | | **Face — identification / verification** | Per 1,000 transactions | Higher rate; matching faces against a person group | | **Spatial Analysis** | Runs on-premises via container | Billed per device/stream per day; not consumed API units | | **Background removal** | Per 1,000 transactions | Segmentation of foreground from background | | **Smart crops** | Per 1,000 transactions | Region-of-interest cropping | **Key pricing mechanics:** - **Free tier (F0):** 5,000 transactions/month for Image Analysis and OCR; sufficient for development. - **Standard (S1) tier:** Per-1,000-transaction pricing with volume tiers (first N million transactions at a higher rate, subsequent at a lower rate). - **Face API:** Requires access approval under Microsoft's Responsible AI policy. Face identification/verification is more expensive than face detection. - **Spatial Analysis container:** Unlike the cloud APIs, Spatial Analysis runs inside a Docker container on an edge device (or VM). The cost model is a flat **per-device-per-day** charge, making it predictable and independent of frame rate or stream volume. - **OCR pricing split:** Images are priced per transaction; PDFs are priced per page. Submitting a 50-page PDF counts as 50 billable units. - **Combining features:** When you call Image Analysis with multiple visual features enabled (e.g. tags + captions + objects), it is still one API call and one transaction. ## What Drives Costs | **Cost driver** | **Why it matters** | | --- | --- | | **Image/document volume** | The primary driver. Each image processed = one transaction regardless of complexity. | | **Feature type selected** | Dense captioning, Face identification, and background removal carry a higher per-transaction rate than basic image tagging. | | **PDF page count** | Submitting multi-page PDFs to the Read OCR API bills per page — not per document. | | **Number of concurrent streams** | Spatial Analysis cost is driven by the number of active containerised deployments, not image volume. | | **Face person group size (identification)** | Large person groups require more compute per identification call; rate stays the same but the use case tends to drive higher volume. | **Cost optimisation levers:** - Only request the visual features you actually need — avoid requesting all available features if only tags are used downstream. - For OCR workloads, split multi-page PDFs and only submit pages that require extraction (e.g. skip cover pages or blank pages). - Cache analysis results keyed on image hash — identical images submitted multiple times each incur a full transaction cost. - Use thumbnail generation or smart crops only when the UI requires it, rather than generating crops for every image by default. - For Spatial Analysis, consolidate multiple camera feeds into the minimum number of containers to minimise the per-device-day charge. ## Azure Cost Data — Meters & Meter Subcategories Azure AI Vision appears under the **Cognitive Services** service family in Cost Management. Resource type: `Microsoft.CognitiveServices/accounts` with kind `ComputerVision`. | **Meter name** | **Meter subcategory** | **What it counts** | | --- | --- | --- | | Computer Vision Transactions | Standard S1 | Image Analysis API calls (tags, objects, captions, brands) | | Computer Vision Dense Caption Transactions | Standard S1 | Dense captioning feature calls | | Computer Vision Background Removal Transactions | Standard S1 | Background removal / segmentation API calls | | Read OCR Transactions | Standard S1 | OCR calls against images (per transaction) | | Read OCR Pages | Standard S1 | OCR pages extracted from PDF documents | | Face Detection Transactions | Standard S1 | Face detection API calls | | Face Identification / Verification Transactions | Standard S1 | Face identification and verification API calls | | Spatial Analysis Devices | Standard S1 | Active Spatial Analysis container deployments billed per device/day | **Meter subcategory notes:** - Computer Vision and Face appear on **separate meters** in billing exports, even though they are hosted under the same `ComputerVision` kind resource — this is because they are conceptually separate APIs despite sharing the account. - OCR has a **dual meter**: image calls generate a Transactions meter row; PDF submissions generate a Pages meter row. Both can appear in the same billing period for the same resource. - The Spatial Analysis meter `Quantity` represents device-days rather than API call counts, which makes it structurally different from all other Vision meters — important to note when building a cost model or anomaly detection rule. - In Cost Management, filtering by `MeterSubCategory = Standard S1` and `ServiceName = Cognitive Services` with `MeterName` containing `Computer Vision` or `Face` is the quickest way to isolate Vision costs from other Cognitive Services on a shared account. - If the Vision resource is shared across teams, resource tagging is the primary cost allocation mechanism — billing data does not carry caller-identity information. --- ## Common Developer Mistakes That Drive Up Costs - **Requesting all visual features when only one is needed.** A single Image Analysis call can request tags, captions, dense captions, objects, brands, colours, and smart crops simultaneously. Each call is still one transaction regardless of which features are enabled — but enabling dense captioning when only tags are consumed downstream adds a higher per-transaction rate unnecessarily. - **Submitting every page of a multi-page PDF to OCR without filtering.** The Read OCR API bills per page for PDFs. Submitting a 100-page document to extract text from 3 relevant pages bills all 100 pages. Pre-filtering to submit only the relevant pages is one of the highest-impact optimisations for document-heavy workloads. - **Re-analysing unchanged images on every pipeline run.** Image processing pipelines that re-submit all images in a blob container on each run (rather than tracking which have been processed) generate repeated transactions for identical images. A product catalogue that adds 10 images per day but reprocesses 10,000 on each daily run pays for 10,000 transactions when only 10 are new. - **Using face identification/verification when face detection is sufficient.** Face identification (matching faces against a person group) is priced at a higher rate than basic face detection. Using identification in scenarios where detection is all that is needed (e.g. checking whether a face is present, not who it is) overpays for the capability used. - **Generating smart crops or thumbnails for every image as a default.** Smart crop and background removal are billed per transaction. Applying these to every image in a catalogue as a blanket default — rather than only for images used in UI contexts where cropping matters — inflates call volume. - **Not caching analysis results by image hash.** Identical images submitted multiple times (e.g. the same product hero image appearing in multiple contexts) each incur a full transaction charge. Caching keyed on the image content hash eliminates duplicate analysis costs. --- ## What Could Make Your Bill Go Up or Down Next Month **Could go UP:** - A new supplier onboarding event brings a large batch of product images to be analysed, generating a one-time spike in Image Analysis transactions. - Dense captioning is enabled globally after a feature change, increasing the per-transaction rate for all image analysis calls. - OCR is extended to process multi-page PDF reports in addition to single-page images, significantly increasing billable pages. - Background removal is added as a default step in the image upload pipeline rather than being triggered on demand. - Face identification is enabled for a new security feature, replacing detection-only calls with higher-rate identification calls. - A pipeline bug removes the processed-image tracking check, causing all images to be re-analysed on every run. **Could go DOWN:** - Image analysis results are cached by content hash; re-submissions of unchanged images are served from cache. - Multi-page PDF submissions are filtered to submit only pages containing relevant content rather than the full document. - Visual features are audited and dense captioning is disabled for workloads where standard captions are sufficient. - Smart crop and background removal are moved from a default pipeline step to an on-demand API call triggered only when the image is selected for a UI context. - A seasonal drop in new product uploads reduces the volume of new images requiring analysis. --- ## Most Common Optimisation Techniques | **Technique** | **Mechanism** | **Typical saving** | | --- | --- | --- | | **Cache analysis results by image content hash** | Store analysis output (tags, captions, extracted text) keyed on image hash; skip API call on cache hit | Eliminates 100% of cost for re-submissions of identical images | | **Filter PDF pages before OCR submission** | Submit only the pages that contain relevant content; pre-process to identify relevant page ranges | Proportional to ratio of relevant pages to total — can be 70–90% on long documents | | **Audit and disable unused visual features** | Remove dense captioning, background removal, or smart crops from the API call if not consumed downstream | Reduces per-transaction rate where premium features are enabled unnecessarily | | **Track processed images to prevent re-processing** | Store processing state (blob metadata, database record) per image; only submit new or changed images | Eliminates reprocessing cost in recurring pipeline runs | | **Use face detection instead of identification where matching is not needed** | Identification rate is higher; detection is appropriate for presence-checking use cases | Per-transaction rate reduction for affected calls | | **Generate smart crops and thumbnails on demand** | Trigger cropping via event (image selected for a feature) rather than for every uploaded image | Proportional to the ratio of images that are actually displayed in crop-requiring contexts |