嵌入式图像信息处理系统的核心算法研究与系统实现

发布时间：2018-05-17 22:29

本文选题：嵌入式系统 + 图像放大　；参考：《浙江大学》2012年博士论文

【摘要】：本文通过对嵌入式图像信息处理系统的核心算法研究,提出了针对图像放大、光学字符识别和并行优化的三方面问题的创新算法。主要的研究内容和创新点包括： 1、提出一种面向低成本嵌入式应用的快速图像在线放大算法。通过将原始图像划分为区域子图,利用查表映射法将其逐一放大形成新区域子图,进而合成放大后的目标图像。以双立方插值为基准动态构建放大表,进一步提出了一种根据领域扩展信息计算可变像素放大表的改进映射方法,有效解决了区域间连接不连续、平滑度低的缺陷。 2、提出一种将曲面表面文字平面化的算法。在光学字符识别方面,针对书籍内缝边缘等在进行扫描和光学字符转化工作中会产生曲面变形的材料,提出一种算法,将曲面表面的文字平面化。首先将文字切割为网格,通过网格的尺寸数据,拟合计算材料变形的表示函数,进而得到变形材料与展平的材料之间的映射关系,将其还原。通过进一步分析,提出一种适合快速计算和实现的不需要拟合函数和积分运算的快速算法。 3、在算法执行优化方面,针对嵌入式平台运算资源有限的特点,提出一种并行优化的硬件级解决方案,描述了一种利用Lagrange乘子约束搜索,能够提高多进程系统性能的高层次综合算法。该算法通过在时间片之间动态调配运算资源来提高各个运算资源的使用率,进而减少总体运算周期,提高性能。该算法可以与现有的高层次综合算法联合使用,并可用现有的大多数工具进行进一步综合。从软件层面上,提出一种利用已有硬件资源,通过调用空闲硬件模块,将部分负载分担的方法,并给出具体应用实例,为已有硬件的软件级优化提出一种思路。 4、根据本文算法,搭建了具体的硬件实现平台,并实现了产品化量产。通过与国际上广泛采用的嵌入式图像处理产品比较,本文算法与硬件实现方法在性能、效果、功耗和成本上都存在优势,具有非常高的实用价值和商业价值。
[Abstract]:Based on the research on the core algorithms of embedded image information processing system, this paper presents an innovative algorithm for image amplification, optical character recognition and parallel optimization. The main research contents and innovations include: 1. A fast image amplification algorithm for low-cost embedded applications is proposed. By dividing the original image into regional subgraphs and using the look-up table mapping method to enlarge them one by one to form a new region sub-map and then to synthesize the amplified target image. Based on bicubic interpolation, an improved mapping method for calculating variable pixel magnification table based on domain expansion information is proposed, which effectively solves the defects of discontinuity and low smoothness between regions. 2. An algorithm is proposed to flatten the text on the surface of a curved surface. In the aspect of optical character recognition, an algorithm is proposed to flatten the text on the surface of the surface, aiming at the materials such as the edge of the inner slit of the book that will produce the deformation of the surface in the process of scanning and transforming the optical characters. Firstly, the text is cut into a grid, and the representation function of the deformation of the material is fitted and calculated by the size data of the mesh, and then the mapping relationship between the deformed material and the flattened material is obtained and restored. Through further analysis, a fast algorithm suitable for fast calculation and implementation without the need of fitting function and integral operation is proposed. 3. In the aspect of algorithm execution optimization, aiming at the limited computing resources of embedded platform, a hardware level solution of parallel optimization is proposed, and a search method based on Lagrange multiplier constraint is described. A high-level synthesis algorithm that can improve the performance of multi-process systems. The algorithm can improve the utilization rate of each computing resource by dynamically allocating the computing resources between time slices, thus reducing the total operation cycle and improving the performance. The algorithm can be used in conjunction with existing high-level synthesis algorithms and can be further synthesized by most existing tools. From the software level, this paper presents a method of sharing part of the load by using the existing hardware resources and calling the free hardware module, and gives a concrete application example to give a way to optimize the software level of the existing hardware. 4. According to the algorithm of this paper, the hardware realization platform is built, and the mass production is realized. Compared with the embedded image processing products which are widely used in the world, the algorithm and hardware implementation method in this paper have advantages in performance, effect, power consumption and cost, so it has very high practical and commercial value.
【学位授予单位】：浙江大学
【学位级别】：博士
【学位授予年份】：2012
【分类号】：TP368.1;TP391.41

【参考文献】