Presents a massively parallel machine for image processing and computer vision, designed to improve the AMP (Autonomous Multiprocessor), a previously designed pure-dataflow-based multiprocessor system for image processing, and to make it faster and more efficient for image processing and computer vision tasks. In the basic design of the AMP there was room for improvement, and, therefore, the authors have recently begun to redesign an improved version of the AMP. The key point of the improvement is to increase the efficiency of execution, especially by optimizing its token matching mechanism, which is indispensable for dataflow-based processors, and the flexibility in its resource management mechanism. The authors first discuss the defects of the previous image processors, then give an overview of the original AMP design. The methodology of its optimization, and the improved system design are presented.