DSP-oriented low level processing for adaptive tuning of video surveillance images

Natural scenes often have a dynamic range (DR, rang of luminance) of up to 6 decades. An image sensor has to resolve this range with approximately 100 values per decade to generate a digital representation containing all details relevant for surveillance [Hecht, Kandel]. Since conventional CCD image sensors have a DR of 2,5 decades (mapped to 8 bit), they are not applicable for surveillance of highly dynamic scenes. Modern image sensors achieve DRs of up to 6 decades, but this range is mapped to more output values, e.g. 4096 values (HDRC, IMS, Germany). However, 12 bit representations are not applicable for most image processing systems, since they are often optimised for 8 bit operations. To benefit from the optimised hardware architecture, high dynamic input images have to be transformed (compressed) into 8 bit representations. Obviously, this compression leads to a decrease of information. We will describe how the transformation from a highly dynamic input image to an 8 bit output image can be performed in real-time and with a negligible loss of information, using a digital signal processor (DSP) for implementation. The advantages of using DSPs are twofold. First, real-time processing of images is often implemented on DSPs. Thus, the transformation can be integrated as a pre-processing step into these systems using the existing hardware. The second advantage is that DSPs give us the flexibility to adapt algorithms to several sensors, hardware settings, and applications.