Discrete wavelet transform (DWT) is an efficient tool for multi-resolution decomposition of images. It has been shown to be very promising due to its high compression ratio and self-similar data structure. Conventionally a 2-D DWT is accomplished by performing two 1-D operations: one along the rows and the other along the columns of an image. Without executing ordered 1-D transforms, we develop a new algorithm to compute a 2-D Haar DWT, the simplest DWT. Two merits of this algorithm are compactness and quickness. The algorithm is implemented with a compact, regular VLSI architecture whose system throughput can be conveniently improved by appropriate parallel/pipeline methods.