Standard Practice for Image Processing to Improve Automated Facial Recognition Search Performance

Abstract

5.1 Images that meet agreed upon international standards (such as ISO/IEC 19794-5: Face Image Data) can normally be submitted to an FRS for searching with little or no operator intervention. 5.2 Manual processing may be beneficial for sub-optimal images (for example, low resolution, heavily compressed, or where the subject’s pose, illumination, or expression are non-neutral). The processing techniques presented in this practice may be applied over an entire image or in localized areas. However, some images may be degraded to the extent that manual processing will not improve the FR search outcome. 5.3 The image processing topics presented in this practice are not intended to override recommendations for maintaining the forensic quality of images intended for a one-to-one comparison. These processes specifically apply to preparing a facial image to submit as a probe into an automated FRS maximizing the likelihood that a potential candidate will be returned in a search result set. 5.4 The goal of any image processing should be to optimize the image for searching by the FRS, not to create an aesthetically pleasing image. An image that looks ‘good’ is not necessarily the same as one optimized for use by an FRS due to image processing done within vendor specific algorithms. 5.4.1 The effect of any manual image processing will vary with different FRS and, in some cases, may degrade performance rather than improve it. 5.4.2 Image processing to the probe image before an FRS search differs from the operational processes performed for a one-to-one manual comparison. 5.4.3 Any decision regarding whether a candidate returned from an FRS search is from the same source as the probe image must be made based on a comparison with the original (unedited) image and NOT the processed image. 5.4.4 Agency-specific or mandated notes and audit trails shall be maintained at all times.