基于深度学习的复杂背景图像分类方法研究

doi:10.16180/j.cnki.issn1007-7820.2020.12.012

摘要/Abstract

摘要：

复杂背景图像受背景干扰后不易被识别。针对这一问题,文中提出了基于前景分割机制的卷积神经网络图像分类方法。采用全卷积神经网络对图像前景区域进行自动分割,通过图像中前景区域周围的最小边界框对其进行定位。对于定位的前景区域,构建卷积神经网络对其进行处理以区分不同的类别,从而实现复杂背景图像的分类。将提出方法在公开数据集中提取的单一背景和复杂背景图像数据集上进行对比实验,并使用迁移学习与数据增强等方法优化模型。实验结果表明,所提方法使用前景区域分割相比于仅分类CNN具有更高的准确度,且复杂背景图像上的准确度提升幅度要远大于单一背景图像。该结果说明引入前景区域分割对于复杂背景图像分类模型准确度的提升具有一定帮助,能够显著前景区域特征并减少背景因素的干扰。

关键词: 深度学习, 卷积神经网络, 图像分割, 前景区域, 复杂背景, 图像识别

Abstract:

To solve the problem that the complex background image cannot be easily recognized by background interference, a convolutional neural network image classification method based on foreground region segmentation mechanism is proposed. The foreground region of image is automatically segmented by using the full convolutional neural network, and is located by the minimum bounding box around it. A convolutional neural network is constructed to distinguish different foreground region categories, thereby realizing the classification of the complex background image. The proposed method is used to perform contrast experiments on simply and complex background image datasets extracted from the public dataset. Some effective methods, such as transfer learning and data augmentation, are used for model optimization. The experimental results show that the proposed method has higher accuracy than only classification CNN on both datasets, and the extent of model accuracy improvement on the complex background image is much larger than that on the simple background image. These results prove that the introduction of foreground region segmentation can improve the accuracy of complex background image classification model, highlight the characteriskics of the foreground region and reduce the interference of background factors.

Key words: deep learning, convolutional neural network, image segmentation, foreground region, complex background, image recognition

中图分类号:

TP391

程俊华,曾国辉,刘瑾. 基于深度学习的复杂背景图像分类方法研究[J]. 电子科技, 2020, 33(12): 59-66.

CHENG Junhua,ZENG Guohui,LIU Jin. Research on Complex Background Image Classification Method Based on Deep Learning[J]. Electronic Science and Technology, 2020, 33(12): 59-66.

图/表 10

图1

图2

图3

图4

图5

图6

表1

图7

表2

表3

参考文献 19

[1]	Bishop C. Pattern recognition and machine learning[M]. Berlin:Springer, 2006.
[2]	Schmidhuber J. Deep learning in neural networks: an overview[J]. Neural Networks, 2015,61(3):85-117. doi: 10.1016/j.neunet.2014.09.003
[3]	Bengio Y, Lecun Y, Hinton G. Deep learning[J]. Nature, 2015,52(1):436-444. doi: 10.1038/052436a0
[4]	周飞燕, 金林鹏, 董军. 卷积神经网络研究综述[J]. 计算机学报, 2017,40(6):1229-1251.
	Zhou Feiyan, Jin Linpeng, Dong Jun. Review of convolutional neural network[J]. Chinese Journal of Computers, 2017,40(6):1229-1251.
[5]	Ahmed A, Hai T. Review of deep convolution neural network in image classification[C]. Piscataway:Intrenational Conference in Redar, Antenna,Microwave,Electronics and Teleccommuntications, 2017.
[6]	龙满生, 欧阳春娟, 刘欢, 等. 基于卷积神经网络与迁移学习的油茶病害图像识别[J]. 农业工程学报, 2018,34(18):194-201.
	Long Mansheng, Ouyang Chunjuan, Liu Huan, et al. Image recognition of Camellia oleifera diseases based on convolutional neural network & transfer learning[J]. Transactions of the Chinese Society of Agricultural Engineering, 2018,34(18):194-201.
[7]	何雪英, 韩忠义, 魏本征. 基于深度学习的乳腺癌病理图像自动分类[J]. 计算机工程与应用, 2018,54(12):121-125.
	He Xueying, Han Zhongyi, Wei Benzheng. Breast cancer histopathological image auto-classification using deep learning[J]. Computer Engineering and Applications, 2018,54(12):121-125.
[8]	王鑫, 李可, 徐明君, 等. 改进的基于深度学习的遥感图像分类方法[J]. 计算机应用, 2019,39(2):1-7.
	Wang Xin, Li Ke, Xu Mingjun, et al. Improved remote sensing image classification method based on deep learning[J]. Journal of Computer Applications, 2019,39(2):1-7.
[9]	Shelhamer E, Long J, Darrell T. Fully convolutional networks for semantic segmentation[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017,39(4):640-651. doi: 10.1109/TPAMI.2016.2572683 pmid: 27244717
[10]	Szegedy C, Loffe S, Vanhoucke V, et al. Inception-v4, Inception-ResNet and the impact of residual connections on learning[J/OL]. (2016-08-23)[2019-03-11] https://arxiv.org/abs/1602.07261.pdf.
[11]	He K M, Zhang X Y, Ren S Q, et al. Deep residual learning for image recognition[C]. Las Vegas:IEEE Conference on Computer Vision and Pattern Recognition, 2016.
[12]	Huang G, Liu Z, Maaten L, et al. Huang G,Liu Z,Maaten L,et al.Densely connected convolutional networks[J/OL]. (2018-01-28)[2019-03-11]. http://arxiv.org/pdf/1608.06993.pdf.
[13]	Krizhevsky A, Sutskever I, Hinton G E. ImageNet classification with deep convolutional neural networks[C]. Pereira:International Conference on Neural Information Processing Systems, 2012.
[14]	Simonyan K, Zisserman A.Very deep convolutional networks for large-scale image recognition[J/OL]. ( (2015-04-10)[2019-03-15]. https://arxiv.org/abs/1409.1556.
[15]	Feng Y Y, Zeng X Y, Yang Y F. Study on the optimization of CNN based on image identification[C]. Wuxi:Distributed Computing and Applications for Business Engineering and Science, 2019.
[16]	Zhang K, Guo Y R, Wang X S, et al. Multiple feature reweight DenseNet for image classification[J]. IEEE Access, 2019(7):9872-9880.
[17]	Mohamed A A, Berg W A, Peng H, et al. A deep learning method for classifying mammographic breast density categories[J]. Medical Physics, 2018,45(1):314-321. doi: 10.1002/mp.12683 pmid: 29159811
[18]	Girshick R. Fast R-CNN[C]. Santiago: Proceeding of IEEE International Conference on Computer Vision,ICCV Press, 2015.
[19]	Zhang C, Li R, Huang Q, et al. Hierarchical deep semantic representation for visual categorization[J]. Neurocomputing, 2017,25(7):88-96.

数据集	数据增强 (有/无)	分割IoU μ±σ/%	检测BIoU μ±σ/%
数据集a	无	82.9±14.3	87.8±12.1
数据集a	有	87.0±12.1	89.5±10.2
数据集b	无	79.8±16.5	84.4±13.6
数据集b	有	85.5±13.2	87.6±11.3

方法	分割(有/无)	分类准确度/%
FCN-CNN(本文)	有	98.13
仅CNN	无	97.85
ResNet-50	无	98.66
Inception-V4	无	99.12
DenseNet-169	无	98.95

方法	分割(有/无)	分类准确度/%
FCN-CNN(本文)	有	97.65
仅CNN	无	93.62
ResNet-50	无	95.87
Inception-V4	无	97.58
DenseNet-169	无	96.35