Hardware Considerations for Illumination-Invariant Image Processing J. W. V. Miller and M. Shridhar Department of Electrical and Computer Engineering The University of Michigan-Dearborn Abstract: Illumination-invariant image processing is an extension of the classical technique of homomorphic filtering using a logarithmic point transformation. In this paper, traditional approaches to illumination-invariant processing are briefly reviewed and then extended using newer image processing techniques. Relevant hardware considerations are also discussed including the number of bits per pixel required for digitization, minimizing the dynamic range of the data for image processing, and camera requirements. Three applications using illumination-invariant processing techniques are also provided. Keywords: illumination invariance, machine vision, quantization, homomorphic filtering, mathematical morphology, gamma correction 1. INTRODUCTION Scene illumination is one of the most important factors that determine success or failure for many imaging and machine vision applications. With appropriate lighting, irrelevant information can be eliminated, contrast of significant image features enhanced, and the sensor signal-to-noise ratio improved. Stringent requirements, however, may be placed on the lighting with regard to the need for very uniform and temporally constant levels of illumination unless suitable low-level processing is performed. Frequent lighting adjustments, tightly regulated power supplies or special feedback circuitry for controlling light level may be needed in these cases which can increase system cost and complexity significantly. Even with ideal illumination, deficiencies in the optics and imaging hardware can introduce apparent variations in scene illumination that reduce the robustness and reliability of vision systems. Optical losses increase significantly from the center of the field to the edge due to the cos4 law and vignetting1. Imagers often exhibit sensitivity variations (shading) as a function of spatial position. While higher-quality system components reduce these problems, the added cost can be excessive. Real-time gray-scale image processing hardware which is generally available in newer vision systems can alleviate the effects of these imperfections economically. Traditionally, linear high-pass filtering prior to thresholding has often been used to reduce the effects of these variations since they are most significant at low spatial frequencies. Because image features of interest such as edges predominate at high spatial frequencies, the effects of illumination variations can be reduced prior to further processing2,3. Morphological processing also has been used to correct for illumination variations4,5. In some applications, a closing operation will generate an image that is a good estimate of background illumination by suppressing dark features at higher spatial frequencies. Calculating the difference between the background image and the original image generates a high-pass-filtered image. High-pass filtering alone, however, does not completely eliminate these effects because most images are approximately a function of the product of the illumination and reflectance properties of a given scene6. An image feature in a dim area of an image will have poorer contrast than the same feature in a bright area as a result of this modulation-like property. A number of techniques have been used to remove the gray-level distortion introduced by this effect including shading correction and homomorphic filtering. The term "illumination invariance" will be used to describe these techniques since they attempt to generate an image in which gray-scale errors associated with illumination, optics and the imager have been eliminated. Errors associated with illumination and optics are presumed to be purely multiplicative while the imager may require both an additive and a multi-plicative correction factor7. Compensation for the multiplicative effects from all three sources can be provided simultaneously after imager offset errors have been removed. Shading correction can directly compensate for these error sources if an "empty" scene is available, that is, a scene which is only a function of the illumination, optics and imager. For purposes of this paper, such a scene is designated as the back-ground. Scene elements of interest, such as objects or image features, comprise the foreground. Hence, if an image of the background can be acquired directly and stored, it can be divided into the foreground image data to compensate for existing nonuniformities1,7 . If imager offset errors are significant, a completely dark scene can be captured and subtracted from theimager data prior to multiplicative correction. Shading correction has been used in a variety of applications including thin-film disk inspection and the evaluation of lumber8,9. Homomorphic filtering represents another approach to illumination-invariant processing10,. Here, gray-scale values of a given image are nonlinearly transformed, the resulting image is linearly filtered, and the inverse of the original nonlinear transformation is performed on the filtered image. Homomorphic filtering using a logarithmic transformation to linearize the multiplicative effect of illumination is a classical and well established image and signal processing technique10. Typical applications include enhancement of photographs containing scenes with wide illumination variations and edge detection11,3. Homomorphic Wiener filtering has also been used for restoration2. In this paper, classical techniques for illumination-invariant processing are reviewed and a number of enhancements are proposed. Homomorphic filtering, for example, has very advantageous properties. However, obtaining linear filters that remove or pass specific scene elements in a given image can be very difficult to implement. Nonlinear filtering operations such as morphological closings can often be generated with the desired characteristics. By logarithmically transforming a given image prior to morphological filtering, illumination-invariance may also be achieved. Implementation issues are also examined here. A detailed discussion regarding the minimization of quantization errors with applications that