Background-Supported Global Feature Response Image Classification Network

doi:10.3778/j.issn.1673-9418.2404027

Abstract

Abstract: The lack of background information support in current image classification methods leads to the limited classification accuracy of the model. Aiming at this problem, a background-supported global feature response image classification network (BGRNet) is proposed. Firstly, based on WRN (wide residual networks) residual networks, a new background-supported activation function BS (background-supported) is proposed, which introduces a background support mechanism through the BS activation function, so that the network can focus on the background information smoothly while focusing on the foreground information of the target. Then, a full-domain feature response module BGR (background-supported global feature response) is proposed, and BGR is embedded into the residual branch to restore the image full domain features, which reduces the loss of feature information due to the convolution operation to a certain extent. Finally, this paper adjusts the internal network structure of the residual block by adjusting the activation function, the forward propagation order of batch normalization and removing Dropout (Dropout Regularization), amplifying the background support role of the BS activation function to the overall network model, and promoting the effective transmission of background information in the network. By introducing the background information support mechanism, BGRNet not only considers the support role of the target foreground information in the process of image classification, but also considers the support role of the background information in the classification process, which effectively improves the network training efficiency while improving the network classification accuracy. Experimental results on FashionMNIST, KMNIST, CIFAR-10, CIFAR-100 and SVHN datasets show that BGRNet significantly improves the classification performance of the baseline model, and compared with the current mainstream methods, BGRNet has higher classification accuracy and stronger generalization performance.

Key words: image classification, background support, full domain feature response, feature restoration, residual network

摘要： 针对目前图像分类方法缺少背景信息支持，导致模型分类精度受限的问题，提出背景支持下的全域特征响应图像分类网络（BGRNet）。该网络以WRN残差网络为基础，提出新的背景支持激活函数BS，通过BS激活函数引入背景支持机制，使网络在关注目标前景信息的同时，也能够平滑地关注背景信息；提出全域特征响应模块（BGR），并将BGR嵌入到残差分支中，对图像全域特征进行还原，在一定程度上减少因卷积操作而产生的特征信息损失；调整残差块内部网络结构，通过调整残差块中激活函数、批量归一化的前向传播顺序并删除Dropout（Dropout Regularization），放大BS激活函数对整体网络模型的背景支持作用，促进背景信息在网络中的有效传递。BGRNet通过引入背景信息支持机制，不仅考虑了目标前景信息在图像分类过程中的支持作用，还考虑了背景信息在分类过程中的支持作用，在提升网络分类精度的同时，有效提高了网络训练效率。在FashionMNIST、KMNIST、CIFAR-10、CIFAR-100、SVHN数据集上的实验结果表明，BGRNet显著提高了基线模型的分类性能，且与当前主流方法相比，BGRNet具有较高的分类准确率和较强的泛化性能。

关键词: 图像分类, 背景支持, 全域特征响应, 特征还原, 残差网络

JIANG Wentao, LI Weida, ZHANG Shengchong. Background-Supported Global Feature Response Image Classification Network[J]. Journal of Frontiers of Computer Science and Technology, 2025, 19(5): 1280-1294.

姜文涛, 李威达, 张晟翀. 背景支持下的全域特征响应图像分类网络[J]. 计算机科学与探索, 2025, 19(5): 1280-1294.

References

[1] LECUN Y, BOTTOU L, BENGIO Y, et al. Gradient-based learning applied to document recognition[J]. Proceedings of the IEEE, 1998, 86(11): 2278-2324.
[2] KRIZHEVSKY A, SUTSKEVER I, HINTON G E. ImageNet classification with deep convolutional neural networks[C]// Proceedings of the 25th International Conference on Neural Information Processing Systems, 2012: 1097-1105.
[3] SIMONYAN K, ZISSERMAN A. Very deep convolutional networks for large-scale image recognition[EB/OL]. [2024-02-17]. https://arxiv.org/abs/1409.1556.
[4] SZEGEDY C, LIU W, JIA Y Q, et al. Going deeper with convolutions[C]//Proceedings of the 2015 IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2015: 1-9.
[5] HE K M, ZHANG X Y, REN S Q, et al. Deep residual learning for image recognition[C]//Proceedings of the 2016 IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2016: 770-778.
[6] HUANG G, LIU Z, VAN DER MAATEN L, et al. Densely connected convolutional networks[C]//Proceedings of the 2017 IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2017: 2261-2269.
[7] DING X, ZHANG X, HAN J, et al. Scaling up your kernels to 31×31: revisiting large kernel design in CNNs[C]//Proceedings of the 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2022: 11963-11975.
[8] DAI Z H, LIU H X, LE Q V, et al. CoAtNet: marrying convolution and attention for all data sizes[EB/OL]. [2024-02-17]. https://arxiv.org/abs/2106.04803.
[9] 付国栋, 黄进, 杨涛, 等. 改进CBAM的轻量级注意力模型[J]. 计算机工程与应用, 2021, 57(20): 150-156.
FU G D, HUANG J, YANG T, et al. Improved lightweight attention model based on CBAM[J]. Computer Engineering and Applications, 2021, 57(20): 150-156.
[10] VASWANI A, SHAZEER N, PARMAR N, et al. Attention is all you need[C]//Advances in Neural Information Processing Systems 30, 2017: 5998-6008.
[11] DOSOVITSKIY A, BEYER L, KOLESNIKOV A, et al. An image is worth 16×16 words: transformers for image recognition at scale[EB/OL]. [2024-02-17]. https://arxiv.org/abs/2010.11929.
[12] 张峰, 黄仕鑫, 花强, 等. 基于Depth-wise卷积和视觉Transformer的图像分类模型[J]. 计算机科学, 2024, 51(2): 196-204.
ZHANG F, HUANG S X, HUA Q, et al. Novel image classification model based on depth-wise convolution neural network and visual transformer[J]. Computer Science, 2024, 51(2): 196-204.
[13] LUO X, WEI L H, WEN L J, et al. Rectifying the shortcut learning of background for few-shot learning[EB/OL]. [2024-03-13]. https://arxiv.org/abs/2107.07746.
[14] 李伟, 黄鹤鸣. 基于双交叉熵的自适应残差卷积图像分类算法[J]. 计算机工程与设计, 2023, 44(12): 3670-3676.
LI W, HUANG H M. Adaptive residual convolution image classification algorithm with dual cross-entropy[J]. Computer Engineering and Design, 2023, 44(12): 3670-3676.
[15] 叶继华, 黎欣, 陈进, 等. MAFDNet: 复杂环境下图像自适应分类新方法[J]. 数据采集与处理, 2023, 38(6): 1392-1405.
YE J H, LI X, CHEN J, et al. MAFDNet: a new method of image adaptive classification in complex environment[J]. Journal of Data Acquisition and Processing, 2023, 38(6): 1392-1405.
[16] ZAGORUYKO S, KOMODAKIS N. Wide residual networks [C]//Proceedings of the 2016 British Machine Vision Conference. Durham: BMVA Press, 2016.
[17] CLEVERT D A, UNTERTHINER T, HOCHREITER S. Fast and accurate deep network learning by exponential linear units (ELUs)[C]//Proceedings of the 4th International Conference on Learning Representations, 2016.
[18] MISRA D. Mish: a self regularized non-monotonic activation function[EB/OL]. [2024-03-13]. https://arxiv.org/abs/1908.08681.
[19] MAAS A L, HANNUN A Y, NG A Y. Rectifier nonlinearities improve neural network acoustic models[C]//Proceedings of the 30th International Conference on Machine Learning. Madison: Omni Press, 2013: 1-9.
[20] NAIR V, HINTON G E. Rectified linear units improve restricted Boltzmann machines[C]//Proceedings of the 27th International Conference on Machine Learning. Madison: Omni Press, 2010: 807-814.
[21] LUU M L, HUANG Z Y, XING E P, et al. Expeditious saliency-guided mix-up through random gradient thresholding[EB/OL]. [2024-03-13]. https://arxiv.org/abs/2212.04875.
[22] QIU X R, ZHU R J, CHOU Y H, et al. Gated attention coding for training high-performance and efficient spiking neural networks[EB/OL]. [2024-03-13]. https://arxiv.org/abs/2308. 06582.
[23] N?KLAND A, EIDNES L H. Training neural networks with local error signals[C]//Proceedings of the 2019 International Conference on Machine Learning, 2019: 4839-4850.
[24] HASSANI A, WALTON S, SHAH N, et al. Escaping the big data paradigm with compact transformers[EB/OL]. [2024-03-13]. https://arxiv.org/abs/2104.05704.
[25] HUANG Z Z, LIANG S W, LIANG M F, et al. DIANet: dense-and-implicit attention network[EB/OL]. [2024-03-13]. https://arxiv.org/abs/1905.10671.
[26] HASANPOUR S H, ROUHANI M, FAYYAZ M, et al. Towards principled design of deep convolutional networks: introducing SimpNet[EB/OL]. [2024-03-13]. https://arxiv.org/abs/1802.06205.
[27] TSENG C H, LEE S J, FENG J N, et al. UPANets: learning from the universal pixel attention neworks[J]. Entropy, 2022, 24(9): 1243.
[28] 郭玉荣, 张珂, 王新胜, 等. 端到端双通道特征重标定DenseNet图像分类[J]. 中国图象图形学报, 2020, 25(3): 486-497.
GUO Y R, ZHANG K, WANG X S, et al. Image classification method based on end-to-end dual feature reweight DenseNet [J]. Journal of Image and Graphics, 2020, 25(3): 486-497.
[29] NAKATA K, NG Y, MIYASHITA D, et al. Revisiting a kNN-based image classification system with high-capacity storage[C]//Proceedings of the 17th European Conference on Computer Vision. Cham: Springer, 2022: 457-474.
[30] LAN H, WANG X H, SHEN H, et al. Couplformer: rethinking vision transformer with coupling attention[C]//Proceedings of the 2023 IEEE/CVF Winter Conference on Applications of Computer Vision. Piscataway: IEEE, 2023: 6464-6473.
[31] BENGIO Y. Learning deep architectures for AI[J]. Foundations and Trends? in Machine Learning, 2009, 2(1): 1-127.