特征提取专用指令集处理器的设计研究与架构优化

As one key technique in computer vision and pattern recognition, feature extraction is a challenging issue. Therefore, this project will make deep research on correctness of feature extraction and efficiency of computing, which will combine the algorithms of feature extraction and design methods of VLSI and computer architecture. For noise problem, spatially-variant adaptive filtering + morphology reconstitution method will used for haze removal, and CNN will used for raindrop and dirt removal. A hierarchical method, which will construct multiple features tree by selecting the optimal feature sets for different tree nodes, is proposed for features fusion to organize and express features. Finally, the design philosophy, which is vision computing instructions recombination, vision processor architecture reconfiguration, and basic computing element reconfiguration, is used to design the feature extraction processor. The design flow will replace function modules integration with instructions recombination. The proposed approach of this project will combine algorithm research and hardware design, which iteratively feedback experiment results and stepwise optimize to realize high performance feature extraction algorithms and processor. The proposed project will contribute to solve some basic and critical problems in computer vision, pattern recognition, and mobile vision device.

作为计算机视觉和模式识别等领域的关键技术，特征提取问题是一项极富挑战的重要课题。本项目拟结合特征提取的相关理论和VLSI以及计算机体系结构的设计方法，围绕特征提取的正确性和计算的高效性问题展开深入研究。针对噪声问题，采用形状可变自适应滤波+形态学重构的方案，实现图像去雾；将CNN用于去雨滴和灰尘，完成特征提取前处理。针对多特征融合，使用层次化方法，选择最佳特征作为子节点，构建多特征树，实现高效特征组织和表达。最后采用视觉计算指令可重组、视觉处理器结构可重构和基本计算单元可重构的设计理念，将传统SoC以功能组装为基础的设计流程转为以功能指令重组为基础，构建特征提取处理器。采用算法研究与硬件设计相结合的方法，通过迭代反馈实验结果，逐步优化，实现高效特征提取算法和特征提取处理器。本项目的研究有助于解决计算机视觉、模式识别和移动视觉设备等领域一些基础且关键性的问题。

作为计算机视觉和模式识别等领域的关键技术，特征提取问题是一项极富挑战的重要课题。本项目结合特征提取的相关理论和VLSI以及计算机体系结构的设计方法，围绕特征提取的正确性和计算的高效性问题展开深入研究。针对图像退化问题，采用形状可变自适应滤波+形态学重构的方案，实现图像去雾；将CNN用于去雨滴和灰尘，完成特征提取前处理。针对多特征融合，使用层次化方法，选择最佳特征作为子节点，构建多特征树，实现高效特征组织和表达。最后采用视觉计算指令可重组、视觉处理器结构可重构和基本计算单元可重构的设计理念，构建特征提取处理器。采用算法研究与硬件设计相结合的方法，通过迭代反馈实验结果，逐步优化，实现了高效特征提取算法和特征提取处理器。本项目的研究有助于解决计算机视觉、模式识别和移动视觉设备等领域一些基础且关键性的问题。