Image processing

The accuracy and reliability of text recognition is highly dependent on the quality of the original image. Aspose.OCR for Python via .NET offers a large number of fully automated and manual image processing filters that enhance an image before it is sent to the OCR engine.

Each filter increases recognition time. The approximate amount of additional time required for preprocessing (as a percentage of the minimum OCR pipeline) is listed in the Performance Impact column.

Filter	Action	Performance impact	Usage scenarios
Skew correction	Automatically straighten images aligned at a slight angle to the horizontal.	12%	Skewed images
Rotation	Manually rotate severely skewed images.	7.5%	Rotated images
Noise removal	Automatically remove dirt, spots, scratches, glare, unwanted gradients, and other noise from photos and scans.	175% extra time; 38% more memory (1)	Photos Old books Newspapers Postcards Documents with stains and dirt
Contrast correction	Automatically adjust the image contrast.	7.5%	Photos Old papers Text on a background
Resizing	Proportionally scale images up / down, or manually define the width and height of the image.	up to 100% (2)	Medication guides Food labels Full-sized photos from modern cameras and smartphones Scanned images at very high DPI
Binarization	Convert images to black and white automatically or manually adjust the criteria that determines whether a pixel is considered black or white.	0.9%	Always used for text detection and most automatic image corrections
Conversion to grayscale	Discard color information from images and leave only shades of gray.	0.5%	Photos Scanned ID cards Full-color scans
Color inversion	Swap image colors so that light areas appear dark and dark areas appear light.	0.25%	White text on black background Advertisements Business cards Screenshots
Dilation	Increase the thickness of characters in an image by adding pixels to the edges of high-contrast objects, such as letters.	3.1%	Receipts Printouts with very thin font
Median filter	Blur noisy images while preserving the edges of high-contrast objects like letters.	6.25%	Photos taken in low light conditions Poor quality printouts Highly compressed JPEG’s
Dewarping	Straighten page curvature and fix camera lens distortion for page photos. This method requires significant resources and time! Consider using it only if the image has geometric distortions preventing accurate recognition.	30-40 seconds; Up to 4 times more memory (3)	Photos of curved pages Ultra wide-angle and fisheye photos Photos from entry-level smartphones

Notes

1. Automatic noise removal uses a specialized neural network that consumes significant computing resources and RAM. Use it with care.
2. Resizing takes between 6% and 100% more time than the minimum processing pipeline, depending on the original image size.
3. Due to the high complexity of the underlying neural network, dewarping is resource- and time-intensive. Actual numbers may vary greatly depending on the performance of the computer and the characteristics of the original image.

Chaining processing filters

Multiple processing filters can be applied to the same image to further improve the recognition quality. The filters are applied one by one in the order they are added to PreprocessingFilter object.

Note that each filter requires additional time and resources on the computer running the application. Do not add extra filters if you are satisfied with the recognition accuracy, especially when developing web applications.

Set binarization threshold

Correct skew

Rotate

Remove noise

Adjust contrast

Scale

Resize

Convert to grayscale

Invert colors

Dilate

Apply median filter

filters = PreprocessingFilter()
filters.add(PreprocessingFilter.threshold({THRESHOLD}))
filters.add(PreprocessingFilter.auto_skew())
filters.add(PreprocessingFilter.rotate({ANGLE}))
filters.add(PreprocessingFilter.auto_denoising())
filters.add(PreprocessingFilter.contrast_correction_filter())
filters.add(PreprocessingFilter.scale({RATIO}))
filters.add(PreprocessingFilter.resize({WIDTH}, {HEIGHT}))
filters.add(PreprocessingFilter.to_grayscale())
filters.add(PreprocessingFilter.invert())
filters.add(PreprocessingFilter.dilate())
filters.add(PreprocessingFilter.median())

Approximate increase of processing time: 0%

Previewing and saving processed images

Aspose.OCR for Python via .NET offers an easy way to access or save processed images using the methods of the ImageProcessing class:

Method	Return value	Description
render()	OcrInput	Applies processing filters to all images in a batch and returns a new batch with processed images. This batch can later be submitted for recognition or used for optimization.
save()	OcrInput	Applies processing filters to all images in batch and saves the resulting images in the specified folder. That method also returns a new batch with processed images, that can be later submitted for the recognition.

You can use these methods to analyze the effectiveness of processing filters, exclude unnecessary filters that consume resources without affecting the result, or show the result of image processing in the user interface.

• Processing filters are applied to all images in the batch, including those without text.
• PDF documents can contain more than one image per page. Therefore, the resulting OcrInput object can contain more images than the number of pages in the document.

# Set processing filters
filters = PreprocessingFilter()
filters.add(PreprocessingFilter.auto_dewarping())
filters.add(PreprocessingFilter.contrast_correction_filter())
# Add all pages from a scanned PDF to OcrInput object and apply processing filters
input = OcrInput(InputType.PDF, filters)
input.add("source.pdf")
# Save processed images from the provided PDF to the "images" folder
ImageProcessing.save(input, "images")