The low-level processing of images is generally a heavy step in vision applications, because the computations, even if very simple, must be iterated for every pixel in the image. Nevertheless, sometimes the processing has a different relevance for different image areas. This fact allows to decrease the number of computations, skipping the pixels which won't produce significant results, and implementing a sort of a multiple ``focus of attention''. This paper presents a hardware extension devoted to the implementation of a data-driven focus of attention on PAPRICA architecture, but can be applied to any SIMD array processor using the same processor virtualization mechanism as PAPRICA. The focus of attention mechanism can be used both to implement different elaborations on different image areas, and to skip the elaboration where it is useless, improving the performances with respect to a traditional architecture.