Scientific publishers often publish older articles (year 2000 and earlier) in scanned form. However, sometimes they seem to have conducted OCR, and added the recovered text as an overlay in order to give the end user a "native PDF" feeling in a sense that it is possible to copy and paste text.
PDFBox differs from other PDF viewers (tested with Adobe Acrobat Reader 7.0, Foxit Reader 3.1, iText 2.1) so that it tries to render both the image part and the textual overlay part, which may produce confusing results.
Actually, there are two separate cases:
*) Page rendering (class org.apache.pdfbox.pdfviewer.PageDrawer): Render the image part and ignore the text part.
*) Text extraction (class org.apache.pdfbox.util.PDFTextStripper): Ignore the image part and work upon the text part.