Photograph a receipt, get the text. Upload a product shot, get keywords for search. Cloud Vision does the hard part; Deluge only has to build a small JSON body and read the answer back.

Before you start

• A Google Cloud project with the Cloud Vision API enabled and billing attached.
• An API key, restricted to the Vision API in the console.
• A publicly reachable image URL, or a Cloud Storage gs:// URI.

Deluge function

// ============================================
// GOOGLE CLOUD VISION — OCR AND LABELLING
// Uses imageUri so nothing has to be base64-encoded.
// ============================================

string Vision.readImage(string imageUrl, string featureType)
{
    API_KEY = "YOUR_GOOGLE_CLOUD_API_KEY";

    source = Map();
    source.put("imageUri",imageUrl);

    image = Map();
    image.put("source",source);

    feature = Map();
    feature.put("type",featureType);        // TEXT_DETECTION or LABEL_DETECTION
    feature.put("maxResults",10);

    features = List();
    features.add(feature);

    request = Map();
    request.put("image",image);
    request.put("features",features);

    requests = List();
    requests.add(request);

    payload = Map();
    payload.put("requests",requests);

    response = invokeurl
    [
        url    : "https://vision.googleapis.com/v1/images:annotate?key=" + API_KEY
        type   : POST
        parameters : payload.toString()
        headers: {"Content-Type":"application/json"}
    ];

    responses = response.getJSON("responses");
    if(responses == null || responses.size() == 0)
    {
        info "Vision error: " + response.toString();
        return "";
    }

    first = responses.get(0);

    // Vision reports per-image failures inside a 200 response
    if(first.getJSON("error") != null)
    {
        info "Vision rejected the image: " + first.getJSON("error").toString();
        return "";
    }

    if(featureType == "TEXT_DETECTION")
    {
        full = first.getJSON("fullTextAnnotation");
        if(full != null)
        {
            return full.getJSON("text");
        }
        return "";
    }
    else
    {
        labels = first.getJSON("labelAnnotations");
        out = "";
        if(labels != null)
        {
            for each l in labels
            {
                if(out != "")
                {
                    out = out + ", ";
                }
                out = out + l.getJSON("description");
            }
        }
        return out;
    }
}

Notes

• Use imageUri, not content. Base64-encoding a photo in Deluge produces a huge string, risks the script timeout, and was the only reason the original needed an external helper. Let Google fetch the image.
• Failures hide inside a 200 response. Vision returns HTTP 200 with a per-image error object when it cannot reach or decode the picture. Without the check above, an unreachable image looks exactly like an image containing no text.
• Each feature per image is one unit. First 1,000 units a month are free, then about $1.50 per 1,000. Asking for TEXT_DETECTION and LABEL_DETECTION on the same photo bills two units. Request only what you will actually use.
• fullTextAnnotation for documents, textAnnotations for words. The first gives you the whole block with layout preserved, which is what you want for a receipt. The second gives every word with bounding boxes, which is what you want when position matters.
• The image must be reachable without a login. A Creator file field behind authentication will not work — push the image somewhere public first, which is one reason to pair this with the Cloudinary script in this section.