【作者】 宋健；

【导师】 李永杰；

【作者基本信息】 电子科技大学 ， 工程硕士（专业学位）， 2022， 硕士

【摘要】 计算机视觉技术深刻地影响了人们的生活,已经在自动驾驶、视频监控、医学成像、智慧交通等生活领域得到应用。然而,在图像采集过程中往往存在一些干扰因素,导致图像存在各种缺陷,降低人类对图像中信息的可解释性和感知能力,图像经过转换、压缩、存储、传输等操作后质量进一步降低,为图片信息的提取和进一步处理带来麻烦,因此对存在缺陷的图像进行增强具有重要意义。现场可编程门阵列作为专用集成电路领域中的一种半定制电路而出现的,拥有强大的计算能力和足够的灵活性,能够实时地处理图像。论文根据生物视觉机制基于现场可编程门阵列实现了两个生物视觉计算模型,可以实时地、高能效地分别增强高动态图像和夜间图像。论文介绍的第一个生物视觉计算模型采用色调映射算法,模拟了视网膜中两种类型的视觉信息处理流,并将视觉信息处理过程划分为四个算法模块,对算法分析后提出了几种使模型在硬件上运行更高效的并行技术,之后对色调映射模型的硬件整体架构以及每个模块的具体实现细节进行了详细描述,最后介绍了三个用于评价增强后图像质量的指标。论文第二个生物视觉计算模型是受生物视觉机制的工作原理启发所建立的夜间图像增强模型。首先介绍了六个算法模块的功能作用以及它们受哪些生物模型的启发,然后介绍了为更高效地实现模型所采用的并行技术,最后对模型的硬件整体架构和各模块实现细节进行了详细描述。最后对两个生物视觉计算模型的各项指标和性能进行了分析,并与相同算法的软件实现结果、其他相关论文的硬件实现结果进行了对比。两个模型的硬件实现可以实时地、高能效地处理1920×1080分辨率的输入图像,输出图像具有较高的信噪比,与软件输出图像具有较高的相似度,同时颜色自然。更多还原

【Abstract】 Computer vision technology has profoundly affected people’s lives and has been applied in life fields such as autonomous driving,video surveillance,medical imaging,and smart transportation.However,there are often some interference factors in the process of image acquisition,which lead to various defects in the image,which reduce the interpretability and perception ability of human beings to the information in the image,and the quality of the image is further reduced after conversion,compression,storage,transmission and other operations.It brings trouble to the extraction and further processing of picture information,so it is of great significance to enhance the image with defects.Field programmable gate arrays emerged as a semi-custom circuit in the field of application-specific integrated circuits,with powerful computing power and sufficient flexibility to process images in real time.According to the biological vision mechanism,the paper implements two biological vision computing models based on field programmable gate arrays,which can enhance high-dynamic images and nighttime images respectively in real time and with high energy efficiency.The first biological visual computing model introduced in this paper adopts the tone mapping algorithm,which simulates two types of visual information processing flow in the retina,and divides the visual information processing process into four algorithm modules.The model runs on the hardware with more efficient parallel technology.After that,the overall hardware architecture of the tone-mapping model and the specific implementation details of each module are described in detail.Finally,three indicators for evaluating the quality of the enhanced image are introduced.The second biological vision computing model of the paper is a nighttime image enhancement model inspired by the working principle of biological vision mechanism.Firstly,the functions of the six algorithm modules and the biological models they are inspired by are introduced.Then,the parallel technology used to make the model more efficient is introduced.Finally,the overall hardware architecture of the model and the implementation details of each module are described in detail.Finally,the indicators and performances of the two biological visual computing models are analyzed,and compared with the software implementation results of the same algorithm and the hardware implementation results of other related papers.The hardware implementation of the two models can process input images with a resolution of 1920 ×1080 in real time and with high energy efficiency,and the output images have high signalto-noise ratio,high similarity with software output images,and natural colors.更多还原