Architectures for Computer Vision

From Algorithm to Chip with Verilog

Hong Jeong

• Verlag: Wiley
• Jahr: 2014
• ISBN:
  9781118659236
• Auflage: 1
• Thema: Computer Science
• Thema: Computer Vision & Pattern Recognition

This book provides comprehensive coverage of 3D vision systems, from vision models and state-of-the-art algorithms to their hardware architectures for implementation on DSPs, FPGA and ASIC chips, and GPUs. It aims to fill the gaps between computer vision algorithms and real-time digital circuit implementations, especially with Verilog HDL design. The organization of this book is vision and hardware module directed, based on Verilog vision modules, 3D vision modules, parallel vision architectures, and Verilog designs for the stereo matching system with various parallel architectures.

• Provides Verilog vision simulators, tailored to the design and testing of general vision chips
• Bridges the differences between C/C++ and HDL to encompass both software realization and chip implementation; includes numerous examples that realize vision algorithms and general vision processing in HDL
• Unique in providing an organized and complete overview of how a real-time 3D vision system-on-chip can be designed
• Focuses on the digital VLSI aspects and implementation of digital signal processing tasks on hardware platforms such as ASICs and FPGAs for 3D vision systems, which have not been comprehensively covered in one single book
• Provides a timely view of the pervasive use of vision systems and the challenges of fusing information from different vision modules
• Accompanying website includes software and HDL code packages to enhance further learning and develop advanced systems
• A solution set and lecture slides are provided on the book's companion website

The book is aimed at graduate students and researchers in computer vision and embedded systems, as well as chip and FPGA designers. Senior undergraduate students specializing in VLSI design or computer vision will also find the book to be helpful in understanding advanced applications.

Part One
Verilog HDL

1
Introduction

This chapter addresses the status of the vision architectures in four major fields: computer architectures, vision algorithms, vision devices, and design methodologies. Computer architecture, which is characterized by serial, parallel, pipelined, and concurrent computation, must be tuned to the underlying computational structures – parallel, iterative, and neighborhood computation – that are used in intermediate computer vision. Vision algorithms, which have evolved from heuristic methods to generic structured algorithms at each level of computer vision from low level to high level, must be investigated in terms of computational structures. The vision devices, ranging from CPUs to very-large-scale integration (VLSI) chips, must be investigated in terms of their flexibility and computational complexity. Finally, the design flow from vision to chip, which is not well-defined, must be defined and delineated using a general methodology.

1.1 Computer Architectures for Vision

Vision architectures are special forms of more general computer architectures. In the early 1970s, a general point of view on computer architecture was to see it as an information flow of data and instructions into a processor (Figure 1.1). Flynn's taxonomy (Flynn 1972) is the most universally accepted method of classifying computer systems. The instruction stream is defined as the sequence of instructions performed by the processing unit. The data stream is defined as the data traffic exchanged between the memory and the processing unit. According to Flynn's classification, the instruction stream and data stream can both be either singular or multiple in nature.

Figure 1.1 Flynn–Johnson taxonomy of computer architectures

Flynn's taxonomy classified architectures into single instruction single data stream (SISD), Single instruction multiple data stream (SIMD), multiple instruction single data stream (MISD), and multiple instruction multiple data stream (MIMD). In this classification system, an SISD machine is the traditional serial architecture where instructions and data are executed serially. This is often referred to as the Von Neumann architecture. An SIMD machine is a parallel computer, where one instruction is executed many times with different data in a synchronized manner. An extreme example is the systolic array (Kung and Leiserson 1980; Kung 1988; Leiserson and Saxe 1991). In an MISD machine, each processing unit operates on the data independently via independent instruction streams. This computational technique is also called pipelining. A set of pipelined vectors is referred to as a superscalar. An MIMD machine is a fully parallel machine where multiple processors execute different instructions and data independently.

This concept can be formalized by a set of state machines, such as the Moore machine or the Mealy machine. Suppose a processing element (PE) in a state Q_k receives date D_k and instruction I_k and generates output O_k, (k = 0, 1, …) according to the state transition T( · ) and output generation H( · ). Then, the SISD machine is modeled by a Mealy machine:

(1.1)

Other machines can be modeled by a set of PEs by combining data and instructions in various ways. As such, an SIMD machine is modeled as a set of identical PEs, operating on different data but controlled by the same instruction set:

(1.2)

where N denotes the number of PEs. The MISD machine is modeled as

(1.3)

where the data input is

and the output is

. The MIMD machine is a set of different machines:

(1.4)

Nowadays, the MIMD category includes a wide variety of different computer types and as a result, taxonomies have been added to the MIMD class. The Flynn–Johnson taxonomy, which is one of many classification methods, proposed a further classification of such machines based on their memory structure (global or distributed) and the mechanism used for communications/synchronization (shared variables or message passing).

A global memory shared variable (GMSV) mach...