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 ›› 2021, Vol. 47 ›› Issue (12): 266-273. doi: 10.19678/j.issn.1000-3428.0060030

• 图形图像处理 • 上一篇    下一篇

基于CornerNet-Saccade的手部分割模型

林竞力, 肖国庆, 张陶, 文鑫   

  1. 西华大学 电气与电子信息学院, 成都 610039
  • 收稿日期:2020-11-17 修回日期:2020-12-17 发布日期:2021-01-05
  • 作者简介:林竞力(1977-),男,副教授、博士,主研方向为图形图像处理、机器学习;肖国庆、张陶、文鑫,硕士研究生。
  • 基金资助:
    国家自然科学基金(61571371);国家自然科学基金青年科学基金项目(61901393);教育部春晖计划合作科研项目(Z201405)。

Hand Segmentation Model Based on CornerNet-Saccade

LIN Jingli, XIAO Guoqing, ZHANG Tao, WEN Xin   

  1. School of Electrical and Information Engineering, Xihua University, Chengdu 610039, China
  • Received:2020-11-17 Revised:2020-12-17 Published:2021-01-05

摘要: 手部分割技术受手部形态、分割背景等因素的影响,分割效率难以提高。在CornerNet-Saccade模型基础上构造一种基于扫视机制的分割模型。通过模拟人眼观察物体时先扫视再仔细观察的行为特征,降低待处理图像的像素数量并在初步判断手部位置后将掩码分支添加到不同尺度特征图中,完成精细分割任务。在此基础上,引入线性瓶颈结构完成模型轻量化操作以降低模型复杂度。实验结果表明,该模型在Egohands数据集上平均交并比高达88.4%,优于RefinNet、U-Net等主流模型,轻量化处理后其平均交并比虽降低了2.2个百分点,但参数量仅为原模型的44.9%。

关键词: 手部分割, 深度学习, CornerNet-Saccade模型, 扫视机制, 轻量化结构

Abstract: The existing hand segmentation technology is limited in the segmentation efficiency due to multiple factors, including various hand shapes and complex segmentation background.To address the problem, this paper optimizes the CornerNet-Saccade model, and on this basis constructs a hand segmentation model using saccade mechanism.This model simulates the action mode of human eyes, which scans a target first, and then observes it carefully.In this way, the model reduces the number of pixels in the to-be-processed image.After preliminary judgment of the hand position, mask branches are added to the feature maps of different scales to complete the fine segmentation task.Moreover, to reduce the complexity of the model, a linear bottleneck structure is introduced to make the model more lightweight.Experimental results show that the mIOU value of the model reaches 88.4% on the Egohands dataset, which is higher than that of mainstream methods such as RefinNet and U-Net.Additionally, the lightweight model further reduces the mIOU value by 2.2% compared with the original model, while its parameters are only 44.9% of the original model.

Key words: hand segmentation, deep learning, CornerNet-Saccade model, saccade mechanism, lightweight structure

中图分类号: