Fixes with the largest effect
Raise effective resolution. Aim for at least 300 DPI on scans, or fill the camera frame with the page. Character height in pixels is the single strongest predictor of accuracy.
Even out the lighting. Diffuse daylight or two lamps at opposing angles removes the shadows and glare that break the ink-versus-paper threshold.
Square up the page. Keep the sensor parallel to the document and rotation within a few degrees.
Fixes worth doing every time
Choose a lossless format. PNG or TIFF preserves character edges; heavy JPEG compression rounds them off.
Crop to the document. Removing backgrounds keeps layout analysis focused on the text.
Declare the language. This tunes the dictionary and language model used in post-processing, which is where marginal characters get resolved.
Fixes for difficult originals
Increase contrast before uploading if the original is faint, but stop short of clipping: pushing the black point too far erases thin strokes entirely.
Split complex pages. Process each column, table or sidebar as its own crop when reading order keeps coming back wrong.
Remove backgrounds and watermarks where you can. Diagonal "COPY" watermarks across text are a reliable source of errors.
Measure rather than guess
Pick one representative page and try two or three capture methods on it. Count the errors in each. Ten minutes of measurement saves hours on a large batch.
Once you find the setup that works for a document type, write it down and reuse it. Consistency across a batch matters more than perfecting any single page.