Azure Computer Vision [Python 3.11] "Input data is not a valid image" when given an image

Question

Hi,

I am working on image extraction from PDFs and using Azure Computer Vision OCR for text extraction in the image. I am using the azure-cognitiveservices-vision-computervision library in Python 3.11, and my Vision resource is a Free tier.

The way my program works is as follows:

• I use the pymupdf library i.e. fitz) to extract images from the pdf:

image_list = page.get_image_info(xrefs=True)
    for img in image_list:
        xref = img["xref"] # get the XREF of the image
        if not xref:
            continue #if the block is not an image, then skip to the next block
         image = fitz.Pixmap(doc,xref)
         #Correct the color space 
         if not image.colorspace.name in (fitz.csGRAY.name, fitz.csRGB.name):
             image = fitz.Pixmap(fitz.csRGB, image) 
		#Save the extracted as a PIL Image
        image = PIL.Image.open(io.BytesIO(image.tobytes())) 

• Perform preprocessing on the image and return the bytes as an BytesIO

buffer = io.BytesIO()
x, y = image.size 
# Clamp the value of the image between 50 and 16000 pixels
new_size = max(50,(min(x,16000))), max(50,min(y,16000))
resized_image = image.resize(new_size)
resized_image.save(buffer, "PNG")
buffer.seek(0)

• Perform OCR using Azure Computer Vision Python API.

result = self.computervision_client.recognize_printed_text_in_stream(buffer)

This works well for most images in PDFs, however I decided to test it on new documents.

One of my test documents contains the following extracted image:

Azure Computer Vision API threw this exception: azure.cognitiveservices.vision.computervision.models._models_py3.ComputerVisionErrorResponseException: (InvalidRequest) Input data is not a valid image.

The image is sent as a PNG file to Azure CV, it has a dimensions of 59 x 93, and it is 58.9KB. I suspected the image may be corrupted or empty, but when saving the contents of the buffer to disk, the image is viewable.

Any help would be appreciated.

Answer 1

@johananmahendran I think the dimensions of this particular image are within range so I would not think that could be an issue. I tried to use the image with the vision studio API and it seems to read the image but it cannot detect any text that can be identified.

In your case, you are using the stable SDK and if it does not detect any text the result should be on similar lines instead of the message you have seen. So, based on this exception message, I think the image stream might be corrupted in the request.

To confirm the same, you can try to pass this image as an URL by uploading it to a public site and then check the response. In this case, you have to just use recognize_printed_text(). I hope this helps!!

If this answers your query, do click Accept Answer and Yes for was this answer helpful. And, if you have any further query do let us know.