Azure Computer Vision in Practice: OCR, Object Detection and Custom Models

What Image Analysis 4.0 gives you

One endpoint, Florence-based: dense captions, OCR (excellent on German umlauts and DIN-format documents), object detection, people detection and smart crops. For SMEs it replaces three separate legacy APIs with a single call.

Shop-floor photos → structured quality data

using Azure;
using Azure.AI.Vision.ImageAnalysis;

var client = new ImageAnalysisClient(
    new Uri(cfg["VISION_ENDPOINT"]),
    new AzureKeyCredential(cfg["VISION_KEY"]));

ImageAnalysisResult result = client.Analyze(
    BinaryData.FromBytes(File.ReadAllBytes("station3_panel.jpg")),
    VisualFeatures.Read | VisualFeatures.DenseCaptions |
    VisualFeatures.Objects);

// 1) serial number from the printed label
string serial = result.Read.Blocks
    .SelectMany(b => b.Lines)
    .Select(l => l.Text)
    .FirstOrDefault(t => t.StartsWith("SN-")) ?? "UNKNOWN";

// 2) anything that looks like damage in the captions
var flags = result.DenseCaptions.Values
    .Where(c => c.Confidence > 0.5 &&
        (c.Text.Contains("scratch") || c.Text.Contains("crack")))
    .Select(c => new { c.Text, Box = c.BoundingBox });

Console.WriteLine($"{serial}: {flags.Count()} potential defects");

Custom models without a data science team

For defect classes generic captions miss, Azure AI Custom Vision trains a detector from a few hundred labeled photos — labelling is an afternoon with the web UI. Export options include ONNX, which leads to the pattern we like most:

Train in Azure, run at the edge

Production lines can't depend on internet latency, and many plants forbid camera images leaving the site. So: train and iterate in Azure, export ONNX, and run inference on an industrial PC at the line — fully offline, with all images deleted after extraction. Azure becomes your tooling environment; the data path stays on premises. That split satisfies both the quality engineer and the works council.