Face detection hardware accelerator using C-based high-level synthesis

• Main Author: Yeap, Han Chien
• Format: Thesis
• Language: English
• Published: 2018
• Subjects: TK Electrical engineering. Electronics Nuclear engineering
• Online Access: http://eprints.utm.my/id/eprint/79563/1/Yeap%2C%20Han%20Chien.pdf; http://eprints.utm.my/id/eprint/79563/
• Institution: Universiti Teknologi Malaysia

Research has shown that Field Programmable Gate Array (FPGA) based implementation of image processing system results in high computational speed and energy efficiency. However, FPGA design has relatively long development time compared to alternative implementation platforms, such as those based on Central Processing Unit, Graphical Processing Unit or Digital Signal Processor. Designing digital hardware at a higher level of abstraction is an effective way to shorten the development time. High-level synthesis (HLS) raises the abstraction level for designing digital circuit and translates a C-based description of the desired design into Hardware Descriptive Language. However, C-based HLS techniques are still lacking some maturity. In particular, existing works on applying C-based HLS to design hardware that accelerates window-based image processing algorithms are generally done in a trial and error manner, and usually results in non-optimal designs. Hence, there is a need for an effective procedure in applying C-based HLS that can lead to an optimized accelerator design. Therefore, the key contribution of this research is to present a systematic C-based HLS technique to be used in the design of hardware that accelerates image processing algorithm. The proposed C-based HLS design procedure is illustrated with a case study of the Sobel filter. The effectiveness of the proposed design technique is demonstrated by the case study of a Viola- Jones face detection accelerator targeted for implementation in FPGA. The proposed face detection hardware applies a pipelined architecture with task-level parallelism that allows concurrent execution on every sub-module. Experimental results show that the resulting accelerator module achieves a speed performance improvement of up to 12 times when compared to that of existing works. Tested on CMU+MIT database, the proposed accelerator achieves high detection accuracy of 88% and 46 false positives. Experimental results also show that the proposed design achieves up to 61 frames per second detection speed. This work demonstrates that the proposed Cbased HLS design methodology is effective for image processing hardware accelerator development.