Journal of Computer Applications ›› 2019, Vol. 39 ›› Issue (11): 3172-3177.DOI: 10.11772/j.issn.1001-9081.2019051140

• The 2019 CCF Conference on Artificial Intelligence (CCFAI2019) • Previous Articles     Next Articles

Data enhancement algorithm based on feature extraction preference and background color correlation

YU Ying, WANG Lewei, ZHANG Yinglong   

  1. College of Software Engineering, East China Jiaotong University, Nanchang Jiangxi 33001, China
  • Received:2019-05-24 Revised:2019-08-17 Online:2019-11-10 Published:2019-09-11
  • Supported by:
    This work is partially supported by the National Natural Science Foundation of China (61563016, 61762036), the Natural Science Foundation of Jiangxi Province (20181BAB202023, 20171BAB202012).

基于特征提取偏好与背景色相关性的数据增强算法

余鹰, 王乐为, 张应龙   

  1. 华东交通大学 软件学院, 南昌 330013
  • 通讯作者: 余鹰
  • 作者简介:余鹰(1979-),女,江西广丰人,副教授,博士,CCF会员,主要研究方向:机器学习、计算机视觉;王乐为(1993-),男,江苏淮安人,硕士研究生,主要研究方向:深度学习、计算机视觉;张应龙(1979-),男,陕西绥德人,副教授,博士,CCF会员,主要研究方向:数据挖掘、网络分析。
  • 基金资助:
    国家自然科学基金资助项目(61563016,61762036);江西省自然科学基金资助项目(20181BAB202023,20171BAB202012)。

Abstract: Deep neural network has powerful feature self-learning ability, which can obtain the granularity features of different levels by multi-layer stepwise feature extraction. However, when the target subject of an image has strong correlation with the background color, the feature extraction will be "lazy", the extracted features are difficult to be discriminated with low abstraction level. To solve this problem, the intrinsic law of feature extraction of deep neural network was studied by experiments. It was found that there was correlation between feature extraction preference and background color of the image. Eliminating this correlation was able to help deep neural network ignore background interference and extract the features of the target subject directly. Therefore, a data enhancement algorithm was proposed and experiments were carried out on the self-built dataset. The experimental results show that the proposed algorithm can reduce the interference of background color on the extraction of target features, reduce over-fitting and improve classification effect.

Key words: feature extraction, data enhancement, deep learning, background color

摘要: 深度神经网络具有强大的特征自学习能力,可以通过多层逐步提取的方式获取不同层次的粒度特征,但当图片目标本体与背景色具有强相关性时,特征提取会存在"惰性",所提取特征的抽象层次较低,判别性不足。针对此问题,通过实验对深度神经网络特征提取的内在规律进行研究,发现特征提取偏好与图片背景色之间具有相关性,消除该相关性可以帮助深度神经网络忽略背景的干扰,直接学习目标本体的特征,由此提出了数据增强算法,并在自主构建的数据集上进行实验。实验结果表明,所提算法可以降低背景色对目标本体特征提取的干扰,减少过拟合,提高分类效果。

关键词: 特征提取, 数据增强, 深度学习, 背景色

CLC Number: