Imae processing, an essential part of any imaging application, is now a mature technology. There's still room for improvement. Here's what's going on.
If you look at a picture of a familiar object and a chunk of the picture is missing, you mentally fill in the missing piece. Image processing tries to make the computer do the same thing. Image processing engines recognize familiar shapes, lines, circles and letters. If a shape is recognized and is incomplete, missing information is added. Extraneous information is removed.
Image processing is important if the scanned images are going to be put through optical mark recognition, forms processing or optical character recognition. When you or I look at a document, we instantly recognize the important information and ignore the noise. Computer recognition engines often can't distinguish noise from the content. Image processing filters out the noise.
Basic image processing includes:
Deskewing. Scans are often crooked. Deskewing straightens documents. Some recognition engines, such as those for barcode recognition, work best when straight lines are perfectly straight and have few jaggies. Deskewing makes the document easier to read on image viewers.
Zoning. Zoning is essential to forms recognition and OCR. Zoning tells the computer what parts of a scan to recognize and what parts to ignore. Zones tell the computer what kind of recognition to use. Zones can also tell the computer what image processing engines should be used on different parts of a scan. If a computer sees a document with only barcode zones and handwriting zones it knows only to use barcode and ICR engines. Not to bother with OMR. This saves time.
Line Completion. Sometimes referred to as character regeneration. This is important for OCR and handwriting recognition. Look at a raw scan of a handwritten page. The strokes that form the handwriting are frequently broken and incomplete. This is because people write with different pressure and different colored inks. Line completion attempts to fill in the missing parts of letters. This is done by having the computer recognize common patterns.
Despeckling. Scans of dirty, creased documents often leave a lot of black spots and noise in the scan. Scanners can accumulate dirt when scanning large numbers of documents. Despeckling removes this unwanted noise. Despeckled documents are easier to compress.
Halftone Removal. Halftones can interfere with character recognition if the characters are printed on top of them. Halftones also make it hard to compress scans. Good image processing software will recognize the distinct patterns of halftone dots and remove them.
Thresholding. Thresholding is the process by which an 8-bit grayscale image is converted to 1-bit black and white. Simple thresholding decides whether a pixel is black or white based on the grayscale value of the pixel alone. Typically, if a pixel grayscale level is less than 128 it becomes black. If it's more than 128, it becomes white.
Sophisticated thresholding doesn't just look at each pixel in isolation. It looks at the area surrounding each pixel to determine if a pixel should be black or white. Sophisticated thresholding gives sharper images and can improve character, handwriting and barcode recognition.
Dithering. Dithering is in some ways the reverse of thresholding. Dithering takes a black and white image and converts it to grayscale. Dithering appears to "blur" the hard edges of black and white scans. Dithering makes black and white text easier to read and gives the illusion of higher resolution. It is especially useful when viewing black and white images on low resolution monitors.
Compression. Small files are easier to store and to transmit across networks than large ones. Software can also parse smaller files faster saving time. During normal operation many image processing applications will reduce file size by removing spots, lines and other hash from images. Image processors can save even more by removing superfluous lines, boxes and pictures from scans. Thresholding is a form of compression. Converting a file from 8-bit grayscale to 1-bit black and white compresses a file by more than 87%.
Color Processing. Color is little used in high volume imaging due to its high memory and processor overhead. Color image processing is confined to image editing software. Color image editing software has long been confined to Macs, Silicon Graphics, and expensive proprietary workstations. With advances on the PC platform such as MMX technology, color image editing software is finally coming to the Windows environment.
Today, most scanning applications include strong image processing. One example is Kofax's (Irvine, CA 714-727-1733) Ascent Capture ($8,000) product. Ascent accomplishes all the scan cleanup operations listed above to improve its capture accuracy.
Trends in Image Processing
Developers are bringing intranet technology to image processing. One interesting application is TMSSequoia's (Stillwater, OK 405-377-0880) ScanFix Java. ScanFix Java is a Java-based user interface intended as a front-end to TMSSequoia's image processing engine. ScanFix Java lets users control image processing from remote computers across an intranet. Users could share a network scanner to scan and process documents from their own computers.
If you're running a full-blown Microsoft shop then you can be comforted to know that similar work is being done in ActiveX. Computer Clearing House (Rochester, NY 716-334-4191) is working to improve ActiveX-based image processing. ActiveX-based image processing can be added to standalone applications as well as Internet applications.
Microsoft has licensed Kodak Digital Science color management software. This will bring color management capabilities to the Windows environment, a feature previously restricted to Mac OS and a few high end Unix platforms.
Color management lets all color display devices, monitors, color printers and color printing presses display the same color despite their differences in technology. Color management is important in publishing and it finds uses in scientific and medical imaging. If you ever prepared a business presentation that worked beautifully on a 20" color monitor but looked dull when printed on paper, then color management software is what you need.
Some new and notable image processing software products are these:
BScan ($2,000) from Image Access (Boca Raton, FL 561-995-8334) is designed as a complete scanning solution. The product has a new barcode scanning engine to improve barcode scanning speed and accuracy.
Cardiff (San Marcos, CA 760-752-5244) is one of the leaders in forms recognition. To make forms recognition work, you need a large and robust suite of image processing tools. Cardiff's image processing is built in to the forms recognition engine so that operation is fast and seamless. The new version of TELEform ($1,500-$5,000) has improved image processing.
Dakota Imaging (Columbia, MD 410-381-3113) is a full-service supplier of document management software. Recently, they added Parascript handwriting recognition to their GroKKer line of forms recognition software (price). Parascript permits free form handwriting recognition. Parascript eliminates the rows of boxes used to control the way forms are filled out.
Dunord (Montreal, Canada 514-284-3123) takes grayscale to 1 bit conversion to new heights with VIR ($2,000). VIR is a toolkit that lets programmers add intelligent grayscale conversion to their image processing systems. Instead of using a simple threshold like 5/4 rounding (which is simple and dumb) VIR determines the grayscale conversion thresholds based on the image being converted. This helps fill in lines and letters without messing up grayscale pictures.
Dunord's Twinstream, part of the VIR toolkit, allows simultaneous grayscale and 1 bit scanning on scanners that support it. The grayscale scan can be zoned and the zones converted to black and white by Dunord's intelligent processing engine. The zones are added back into the 1-bit scan. Pictures look better after grayscale processing than a simple 1-bit scan. Barcodes taken from the grayscale scan are easier to read. This speeds up throughput and improves accuracy.
Document scanning and capture are almost synonymous with Kofax (Irvine, CA 714-727-1733). Just about every product they make incorporates some type of image processing. Of particular interest is Image Controls ($1,000-$5,000), a scanning toolkit, and Ascent Capture ($8,000), Kofax's shrink-wrapped scanning product.
If you want government strength image processing, you want Optimum Solutions (Rockville Centre, NY 516-763-6510) FAQSS. FAQSS is a popular image processing solution with market research firms who need optical mark recognition. On the strength of its patented software, FAQSS' image processing and forms recognition engine is among those selected by the Census Bureau for the next Census two years from now.
