In this paper, we introduce a novel technique for memory compression of the Hough transform. Our approach is to partition an image into many sub-images, and there is no need to iterate the line detection with different levels of hierarchical line finding. On the average, our algorithm only requires a few seconds for the extraction of global edges for images with 384*256 pixels using a 33 MHz 486 machine. Besides the reduction of processing time as comparing with the conventional approach, a 50% reduction of the conventional memory requirement can also be achieved for partitioning the image into 16 sub-images. The novel Hough domain consists of two parts. They are (1) the Hough voting space to locate the parameters of straight edges in the image, and (2) the Region-Bit-Map for each sub-image to identify the corresponding peaks. A careful study of the effect for the deviation of ρ which is the normal distance of an edge pixel, has also been made. A range for the increment of ρ is suggested, which can substantially improve the accuracy of the whole approach.