SDSDK - How does the static compare work?
The Signature Image Validation (SIVAL) engine is built as a neural network. The engine compares two signatures and returns a match score.
SIVAL works on black & white TIFFs images. However, signature images as such cannot be fed to the neural network for comparison. Instead, several features relevant to handwriting are extracted from the signature. These include both scalar parameters as overall signature complexity, aspect ratio, etc., and vector parameters collected along with different directions.
When SIVAL looks at the signature, it takes the signature image, computes the list of parameters from it, and then compares the parameters with a reference set. SIVAL has been trained to respond for most of the items with an acceptance rate close to 100, or 0. This needs to be considered when setting the system acceptance threshold.
The scope for which SIVAL was designed is to recognize the signature on a document given a set of reference signatures to choose from. SIVAL’s main task, which it has been trained for, is to identify which signer from a set of reference signatures can be found on the document. The engine will return a confidence value for the comparison. It is up to the application to call the engine to set a threshold above which signature pairs are treated as matching.
There are some important factors that affect the engine decision, mainly reference data, image quality, size and resolution, and system settings.
The quality of the reference data is important. If it does not match system requirements, the results will not be accurate. Since reference data is used repetitively, it is a good idea to verify the image quality
(signature correctly cut out, well cleaned, high resolution, no lines, stamps, etc.).
Bad quality or missing reference data is the most common cause of a poor acceptance rate.
As mentioned, image quality is very important. All artifacts remaining in the signature image that is presented for verification will be treated by SIVAL as part of the signature. The cleaning capabilities built-in the toolkit can be used if the signature image needs pre-processing. The toolkit provides functionality to:
- Remove background dirt
- Remove lines from the image
- Remove printed text from the image (text items overlapping signatures will not be cleaned)
The cleaning function will not remove any part that is connected to the signature itself. Thus, if the signature is on printed text or a stamp has been applied to the signature, the parts or letters that connect to signature lines will not be effectively removed. Furthermore, handwritten text can’t be distinguished from the signature during cleaning.
A good document layout that keeps additional data separated from the area where the signature will be placed can also greatly increase verification quality.
Another factor affecting the quality of the verification results is the scanning quality. A good contrast level is essential to keep image noise out of the picture.
Verification quality also depends on image resolution. SIVAL is designed to work on images with a resolution of 150 dpi or better. In a system that has not been optimized, quality increases with higher resolutions images. Using 200 or 300 dpi images is a good compromise between quality and size.
In case verification problems occur, please verify the following:
- Look at a representative sample of the images you consider falsely accepted or rejected.
- Check that reference data is available and has a good reference image.
- Verify image quality:
- Image resolution is high enough (150 dpi or more) and has been maintained throughout image processing.
- The image type is supported (SIVAL itself works on black & white TIFFs).
- The contrast of the image is good.
- No noise or additional objects are present (text, part of the date, stamps, etc.). If unsure that cleaning is applied, check cleaning settings and dump images after cleaning.
- The snippet area contains the full signature (and nothing but the signature).
Level of Complexity
Add any references to other internal or external articles