A no-reference error-tolerability test technique for videos via edge and extreme-value checking and its hardware implementation

摘要

Approximate computing has been shown to be a promising manner to enhance circuit performance via acceptable quality degradation. This manner is applicable to many multimedia applications such as videos. In order to carry out a dynamic approximate computing scheme, on-line quality evaluation/monitoring will be needed. However, to the best of our knowledge, on-line error-tolerability evaluation of videos is seldom studied in the literature. To achieve this, a no-reference manner will be helpful, which means that no reference (golden) videos are needed for comparison with the videos under test. Implementation of on-line test procedures can thus be greatly simplified. In this paper we investigate how to achieve no-reference error-tolerability evaluation for videos with fixed or varied background. For videos with fixed background a no-reference error-tolerability test technique is proposed. By well exploiting some simple attributes, the proposed technique can accurately determine the acceptability for 81,412 erroneous videos with > 90% accuracy. As a comparison, the related previous work can only achieve about 80% accuracy. In addition, our attribute acquirement process requires only 33% of the computation time for the previous work. As for videos with varied background, we also discuss possible solutions. Hardware implementation of the proposed test technique is also addressed. The results show that the hardware contains about 13,702 logic gates, which is 8.11% of an open-source baseline H.264 video decoder. The proposed hardware operates at 20 MHz frequency, which is sufficient to support real-time error-tolerability testing of videos.