面向实时图像识别的物体基本形状特征提取的高速化研究

Extracting basic shape features of objects in a binary image, such as area, perimeter, circularity ratio, bounding box, geometric center, Euler number, contour etc. is indispensable to almost all image recognition systems. For pattern recognition in real-time image processing systems such as online detection, robot vision, auto-driving system, and optical imaging guidance weapon systems, where dozens or hundreds of images should be processed in a second, in order to have more time for complicated object recognition processing, it is necessary to extract basic shape features of objects in images as faster as possible. In this project, we will try to speed up the process of extracting basic shape features of objects in a binary image by three ways: First, we will propose more efficient connected component labeling algorithm and Euler number computing algorithm by reducing the times for checking neighbor pixels; Second, we will propose an integrated algorithm for extracting various shape features of objects in a binary image simultaneously by use of the related intermediate information obtained during the process of connected component labeling; Lastly, based on the algorithms mentioned above, we will propose the related algorithms suitable for hardware implementation and parallel processing, and implement them by use of FPGA. By the achievements of this project, the processing time for extracting shape features of objects in a binary image will decrease a lot, and the performance of real-time image recognition systems can be improved greatly. Thus, our research can make some contributions for the development of computer science and artificial intelligence.

提取二值图像中物体的基本形状特征（面积、周长、圆形度、边界框、形心、欧拉数、轮廓等）是大多数图像识别系统中不可或缺的底层处理。在每秒钟要处理几十上百张图像的高速在线检测、机器人视觉、自动驾驶及光学成像制导等实时图像识别系统中，提取物体的基本形状特征的速度直接关系到系统的整体性能。本项目将从三个方面实现物体基本形状特征提取的高速化：通过减少对邻接像素的检测次数来加快相关处理，提出比现有方法更有效的连通域标记算法及欧拉数计算算法；开创性地提出通过利用在对图像进行连通域标记处理时保存的相关信息来快速提取图像中各个物体的各种形状特征的综合算法；对上述提出的算法进行分析，提出适合硬件安装及并行处理的相应算法并用FPGA加以实现。本项目的研究成果将大大缩短从二值图像中提取物体基本形状特征的处理时间，为提高以机器视觉为代表的实时图像识别系统的性能提供支持，为推动计算机科学和人工智能的进步做出贡献。

随着计算机技术的发展，图像分析与识别成为国内外众多学者研究的热点，在工业、农业、医疗、环境和军事等领域具有广泛的应用前景。提取二值图像中物体的基本形状特征（面积、周长、圆形度、边界框、形心、欧拉数、轮廓等）是大多数图像识别系统中不可或缺的底层处理。在每秒钟要处理几十上百张图像的高速在线检测、机器人视觉、自动驾驶及光学成像制导等实时图像识别系统中，提取物体的基本形状特征的速度直接关系到系统的整体性能，为此，本项目开展了针对实时图像识别的物体基本形状特征提取高速化处理问题的关键理论和方法的系统研究,通过利用处理过的像素信息、同时处理多行等手段减少对邻接像素的检测次数达到加快相关处理的目的，提出比现有方法更有效的连通域标记算法及欧拉数计算算法；提出了在进行连通域标记处理的同时计算欧拉数的方法；并基于上述算法提出了提取图像中各个物体的各种形状特征的快速算法；对上述提出的算法进行分析，提出适合硬件安装及并行处理的相应算法以及用FPGA加以实现等解决方案，并在多种不同的图像测试数据集上测试了所有所提算法的正确性和有效性。本项目的开展取得一系列创新性成果，大大缩短从二值图像中提取物体基本形状特征的处理时间，为提高以机器视觉为代表的实时图像识别系统的性能提供支持，为推动计算机科学和人工智能的进步做出了贡献。