ACIP - Approximate Computing for Professional Image Processing

The recent advancements in embedded computing devices have been lagging behind the increasing demand for low power consumption and high performance in professional image processing applications. Approximate computing provides a new design paradigm for efficient system design to fill this gap by exploiting the resilience of applications to inaccuracy in their computations and effectively trading off the application quality for hardware resource savings such as power and area as well as performance improvements. As a result, fewer resources are used, more functions can be implemented, and the energy efficiency of the calculations is improved.

Many approximation methods for FPGA-based systems have been proposed in previous years and their suitability has been demonstrated well in isolation. The ACIP project focuses on analyzing the effectiveness of combining multiple single-purpose methods and implementing them in various image processing applications such as digital motion picture cameras and laser scanners.

Research Focus

This research project aims to develop an approximation toolbox called ApproxTB for FPGA-based image processing that can be used to combine multiple approximations in an application and is capable of assessing the trade-off between the loss of application quality and hardware resource benefits. This primarily includes the investigation of

  • Various approximate computing methods relevant for FPGA-based image processing
  • Hardware resource models and suitable error metrics for application quality estimation
  • Optimization approaches that can be used to globally optimize the parameters exposed by the combination of multiple approximations in an application and perform the trade-off analysis

Developments and Status of the Project

The investigations and experimentation for the development of AppoxTB are currently progressing. The general idea is to represent an FPGA application with a data flow graph (DFG) where the nodes represent components or operations of the application. Each node in the DFG can be replaced by one or more components from a library of parameterizable approximated components in the ApproxTB to form an approximated design. This library includes hardware implementations, behavioral models for software simulation, and the basis of the resource models for each approximated component. The exposed parameters from the approximated design make design space complex, and fitness values of design points are estimated using resource models and application-specific quality metrics. A global design space exploration using multi-objective search heuristics is applied to form a Pareto front for effective trade-off analysis between the competing objectives.

The initial developments are

  • Selection of approximation methods that expose parameters for ApproxTB such as approximate adders, sparse look-up tables, bitwidth truncation
  • Sourcing of power and resource models for ApproxTB, targeting an Intel Arria 10 FPGA
  • Develop a quality evaluation system that can employ a wide range of reference metrics to cope with a wide variety of applications
  • Selection of nondominated sorting genetic algorithm (NSGA) as a suitable optimization approach for effective trade-off analysis
  • Experimentation with commonly used image processing applications, for example, a display rendering application or a color space conversion system, for the demonstration of ApproxTB and the proposed approximate computing concept
  • Development of an FPGA emulator system that can be used for prototyping, model verification and to accelerate the optimization process

Funding and Cooperations

The ACIP project is done in close cooperation with Arnold & Richter Cine Technik (ARRI), SmartRay GmbH, and Technical University of Munich (TUM). This project is funded by StMWi Bayern (IuK program).

Offered Student Works


Are you interested in contributing to ACIP project? If you don't find an interesting topic listed here, sometimes there is also the possibility to define a topic matching your specific interests. If you have questions concerning ACIP project and student works please contact Manu Manuel.


  • 7/7
    Manu Manuel, Benjamin Hien, Simon Conrady, Arne Kreddig, Nguyen Anh Vu Doan, and Walter Stechele: Region of Interest Based Non-dominated Sorting Genetic Algorithm-II: An Invite and Conquer Approach. The Genetic and Evolutionary Computation Conference (GECCO), 2022 mehr… BibTeX Volltext ( DOI )
  • 6/7
    Simon Conrady, Arne Kreddig, Manu Manuel, Nguyen Anh Vu Doan, Walter Stechele: Model-based design space exploration for FPGA-based image processing applications employing parameterizable approximations. Microprocessors and Microsystems, 2021, 104386 mehr… BibTeX Volltext ( DOI )
  • 5/7
    Manu Manuel, Arne Kreddig, Simon Conrady, Nguyen Anh Vu Doan, Walter Stechele: Region of Interest-Based Parameter Optimization for Approximate Image Processing on FPGAs. International Journal of Networking and Computing 11 (2), 2021, 438-462 mehr… BibTeX Volltext ( DOI )
  • 4/7
    Arne Kreddig, Simon Conrady, Manu Manuel, Walter Stechele: A Framework for Hardware-Accelerated Design Space Exploration for Approximate Computing on FPGA. 2021 24th Euromicro Conference on Digital System Design (DSD), IEEE, 2021 mehr… BibTeX Volltext ( DOI )
  • 3/7
    Nguyen Anh Vu Doan, Manu Manuel, Simon Conrady, Arne Kreddig, Walter Stechele: Parameter Optimization of Approximate Image Processing Algorithms in FPGAs. 2020 Eighth International Symposium on Computing and Networking (CANDARW), 2020 mehr… BibTeX Volltext ( DOI )
  • 2/7
    Manu Manuel, Arne Kreddig, Simon Conrady, Nguyen Anh Vu Doan, Walter Stechele: Model-Based Design Space Exploration for Approximate Image Processing on FPGA. 2020 IEEE Nordic Circuits and Systems Conference (NorCAS), 2020 mehr… BibTeX Volltext ( DOI )
  • 1/7
    Simon Conrady, Manu Manuel, Arne Kreddig, Walter Stechele: LCS-Based Automatic Configuration of Approximate Computing Parameters for FPGA System Designs. Proceedings of the Genetic and Evolutionary Computation Conference Companion (GECCO '19), 2019, 1271 -- 1279 mehr… BibTeX Volltext ( DOI )