Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Zehua Dong has a total of 5 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • PEPPR-DWS on FPGA: Elevating Universal Parallelism and Precision Through Pulse-Enhanced Push-Relabel and Diffusion Wave Search

    by Zehua Dong, Boyu Zhang, Yucheng Jiang, Yu Yu, Han Li, Songping Mai

    The push-relabel algorithm is recognized as one of the efficient algorithms in the field of graph cut, finding widespread applications in computer vision. While its pixel-level parallel implementations are prevalent, existing methods predominantly rely on checkerboard scheduling, imposing inherent constraints on neighborhood size, limited to four. This limitation compromises both algorithm precision and efficiency, hindering real-time and high-precision applications. To address these issues, this article introduces a novel approach to accelerate push-relabel algorithm implementation on FPGA in a more universal and efficient manner, supporting variable-sized image block operations. First, by introducing the deferred update strategy, we realize the pulse-enhanced parallel push-relabel (PEPPR) algorithm to address data contention and conflict in parallel processing. Second, the simultaneous weighted push method is proposed, further enhancing parallel operations. Lastly, we introduce the efficient diffusion wave search (DWS) algorithm to expedite algorithm convergence and reduce redundancy. While achieving a modest $1.7\times $ acceleration compared to state-of-the-art implementations, the proposed algorithm (PEPPR-DWS) successfully overcomes the inherent limitations of checkerboard scheduling in full pixel-level parallelism. In the test based on Middlebury benchmark V3, the proposed 8-neighborhood implementation exhibits a reduction of error rate by over 1% compared to the typical 4-neighborhood implementation. It provides a versatile and efficient solution for high-precision and real-time applications, holding substantial potential for practical applications.


  • A Configurable Local-Expansion-Move Accelerator for Accurate Stereo Matching

    by Yucheng Jiang, Yu Yu, Han Li, Zehua Dong, Songping Mai

    High-precision stereo matching implementation requires substantial computational resources. To achieve precise real-time depth estimation on embedded devices, we propose a hardware accelerator for local-expansion-move (LE) that can be utilized for accurate disparity extraction. By inferring per-pixel 3D plane labels, it is able to obtain smooth disparities. With the efficient multi-level pipeline design and sophisticated memory structures, we have implemented two key techniques, localization and spatial propagation, which significantly enhance the efficiency and precision of the inference. Compared to the software implementations of LE on CPU and GPU, the accelerator exhibits speedups of 524x and 27.5x respectively. It achieves state-of-the-art matching accuracy in embedded systems by integrating with a lightweight neural network, while maintaining quasi-real-time efficiency. Moreover, the accelerator is implemented using high-level synthesis, which allows for easy modification of accelerator parameters based on the image resolutions and performance requirements in different applications.


  • Open Access

    Robust Cost Volume Generation Method for Dense Stereo Matching in Endoscopic Scenarios

    by Yucheng Jiang, Zehua Dong, Songping Mai

    Stereo matching in binocular endoscopic scenarios is difficult due to the radiometric distortion caused by restricted light conditions. Traditional matching algorithms suffer from poor performance in challenging areas, while deep learning ones are limited by their generalizability and complexity. We introduce a non-deep learning cost volume generation method whose performance is close to a deep learning algorithm, but with far less computation. To deal with the radiometric distortion problem, the initial cost volume is constructed using two radiometric invariant cost metrics, the histogram of gradient angle and amplitude descriptors. Then we propose a new cross-scale propagation framework to improve the matching reliability in small homogenous regions without increasing the running time. The experimental results on the Middlebury Version 3 Benchmark show that the performance of the combination of our method and Local-Expansion, an optimization algorithm, ranks top among non-deep learning algorithms. Other quantitative experimental results on a surgical endoscopic dataset and our binocular endoscope show that the accuracy of the proposed algorithm is at the millimeter level which is comparable to the accuracy of deep learning algorithms. In addition, our method is 65 times faster than its deep learning counterpart in terms of cost volume generation.