作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2022, Vol. 48 ›› Issue (5): 258-262,271. doi: 10.19678/j.issn.1000-3428.0061392

• 图形图像处理 • 上一篇    下一篇

基于多尺度分割的图像识别残差网络研究

袁单飞1,2, 陈慈发1,2, 董方敏1   

  1. 1. 三峡大学 计算机与信息学院, 湖北 宜昌 443000;
    2. 湖北省建筑质量检测装备工程技术研究中心, 湖北 宜昌 443000
  • 收稿日期:2021-04-24 修回日期:2021-05-31 发布日期:2022-05-10
  • 作者简介:袁单飞(1993—),男,硕士研究生,主研方向为计算机视觉、人工智能、嵌入式系统;陈慈发(通信作者),教授、研究员;董方敏,教授、博士生导师。
  • 基金资助:
    国家自然科学基金新疆联合基金重点项目(U1703261)。

Research on Residual Network of Image Recognition Based on Multiscale Split

YUAN Danfei1,2, CHEN Cifa1,2, DONG Fangmin1   

  1. 1. College of Computer and Information Technology, China Three Gorges University, Yichang, Hubei 443000, China;
    2. Hubei Province Engineering Technology Research Center for Construction Quality Testing Equipment, Yichang, Hubei 443000, China
  • Received:2021-04-24 Revised:2021-05-31 Published:2022-05-10

摘要: 深度卷积神经网络能够解决复杂的计算机视觉问题,被广泛应用于图像识别任务中。在基于深度卷积神经网络的图像识别过程中,增加网络的深度和宽度能够产生丰富的特征信息,使用多尺度分割方法能够有效减少冗余的特征信息。然而,增加网络的深度和进行多尺度分割都会影响识别速度。如何在保证精度的同时提高识别速度,成为设计高效网络的关键问题。通过增加网络宽度的方法对ResNet残差网络进行改进,在保证精度的基础上提升识别速度。使用ResNet-D中的残差结构并减少网络长度,得到长度只有7层的残差网络,同时对HS-ResNet中的多尺度分割方法进行优化,只保留最后一次连接合并操作,得到图像识别残差网络SSRNet。在CIFAR 10和CIFAR 100数据集上的实验结果显示,SSRNet速度最高较ResNet网络提升7倍多,同时错误率最高下降8.81%,表明缩短网络长度可大幅加快图像识别速度,同时结合多尺度分割方法能够有效提升识别精度。

关键词: 多尺度分割, 残差网络, 卷积神经网络, 图像识别, 图像分类

Abstract: The emergence of deep Convolutional Neural Network(CNN) has contributed significantly to solving complex computer vision problems, and they have been widely used in image recognition tasks.Designing an efficient network has become the key to improving the performance of deep CNN.In the process of image recognition based on a deep CNN, increasing the depth and width of the network can produce rich feature information, whereas the use of a multi-scale segmentation method can effectively reduce redundant feature information.However, increasing the depth of the network in multi-scale segmentation affects the recognition speed.Thus, improving recognition speed while ensuring accuracy has become an important goal in designing efficient networks.To solve this problem, the network width is increased using ResNet to improve the recognition speed and ensure accuracy.Using the residual structure in ResNet-D and reducing the network length, a residual network with only seven layers is obtained.Concurrently, the multi-scale segmentation method in HS-ResNet is optimized, and only the last connection and merging operation are retained to obtain SSRNet.The experimental results on the CIFAR 10 and CIFAR 100 datasets show that the maximum speed of SSRNet is more than seven times higher than that of ResNet, and the error rate can be reduced by 8.81%.This demonstrates that shortening the length of the network can significantly accelerate the speed of image recognition, whereby the recognition accuracy is effectively improved in combination with the multi-scale segmentation method.

Key words: multiscale split, residual network, Convolutional Neural Network(CNN), image recognition, image classification

中图分类号: