一种图像形状的编码表达方法及其应用  

An image shape coding expression method and its application

作　　者：苗军 许少武 卿来云 乔元华[3] 邹柏贤[4] MIAO Jun;XU Shaowu;QING Laiyun;QIAO Yuanhua;ZOU Baixian(Computer School,Beijing Key Laboratory of Internet Culture and Digital Dissemination Research,Beijing Information Science&Technology University,Beijing 100101,China;School of Computer Science and Technology,University of Chinese Academy of Sciences,Beijing 100049,China;College of Applied Sciences,Beijing University of Technology,Beijing 100124,China;College of Applied Arts and Science,Beijing Union University,Beijing 100191,China)

机构地区：[1]北京信息科技大学计算机学院,网络文化与数字传播北京市重点实验室,北京100101 [2]中国科学院大学计算机科学与技术学院,北京100049 [3]北京工业大学数理学院,北京100124 [4]北京联合大学应用文理学院,北京100191

出　　处：《北京信息科技大学学报（自然科学版）》2020年第6期1-7,共7页Journal of Beijing Information Science and Technology University

基　　金：国家自然科学基金项目(61872333);北京市自然科学基金项目(4202025);北京市教委科技计划项目(KM201911232003);北京市未来芯片技术高精尖创新中心科研基金(KYJJ2018004)。

摘　　要：针对一般卷积结构无法直接提取图像的高级语义特征的问题,提出了通过编码的方式获取形状这一图像全局结构特征的方法——形状编码。形状编码包含两个步骤:第一步是将原图像转换为由显著像素点和非显著像素点组成的二值特征图;第二步是基于二值特征图中显著点对的空间位置关系进行编码。编码的结果是表征原图像的形状特征的形状编码图,可用于替代原始图像送入卷积神经网络中作为学习对象。在形状编码方法的基础上提出了两种改进编码方法,分别是动态形状编码和分块形状编码。实验证明,同时将形状编码图和原始图像送入卷积神经网络进行学习,相比只使用原始图像时可以获得更高的识别准确率。For the problem that the general convolutional structure cannot directly extract the high-level semantic features of the image,a method for obtaining the global structural features of the image by encoding is proposed.This method is called shape encoding,which consists of two steps.The first step is to convert the original image into the binary feature map composed of salient pixels and insignificant pixels,and the second step is to encode the map based on the spatial position relationship of salient point pairs in the map.The result of the encoding is the shape coding map that represents the shape feature of the original image,which can be used to replace the original image and be sent to the convolutional neural network as a learning object.Two improved coding methods based on the shape encoding,namely dynamic shape encoding and partitioning shape encoding are proposed.It is proved through experiments that sending the shape coding map and the original image into the convolutional neural network for learning at the same time can achieve higher recognition accuracy than that when only using the original image.

关 键 词：形状编码 卷积神经网络 手写数字识别 图像识别 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]