跨通道细粒度特征融合的矿石图像分类算法

doi:10.3778/j.issn.1002-8331.2401-0188

摘要/Abstract

摘要： 为解决深度学习算法在处理细粒度纹理特征的矿石图像时准确率低、计算资源需求大且难以在移动端部署的问题，提出一种跨通道细粒度特征融合的轻量级矿石图像分类算法。通过交替使用CNN与Transformer构建混合网络，以有效提取图像局部与全局信息；引入跨通道细粒度特征融合模块作为特征融合器，采用通道分组和随机通道混洗的融合策略，增强矿石纹理信息的获取能力和保持细粒度特征的多样性；利用多尺度轻量化自注意力模块降低模型参数，增强对不同尺度和空间位置的感知，确保训练的稳定性并避免过度拟合低级特征；构建高效坐标注意力模块作为细粒度特征提取器，实现轻量化和高效率的特征提取。所提算法在Kaggle平台的Mineral Photos和Petrology Thin Section Data两个公开矿石图像数据集上分别取得了95.78%和94.77%的分类准确率，相较于其他9种轻量级分类网络，如ShuffleNetV2、MobileNetV3、RegNet、ConvNeXtV2、LeViT、EdgeViTs、AFFNeT、EdgeNeXt和MViTV2，所提算法具有更少的参数（1.27?MB）、更低的计算量（269?MFLOPs）和更快的分类速度（219?FPS）。

关键词: 矿石图像分类, 卷积神经网络（CNN）, Transformer, 跨通道特征融合, 注意力机制

Abstract: In order to solve the problems that the deep learning algorithm has low accuracy in processing ore images with fine-grained texture features, large computing resource requirements and difficulty to deploy on the mobile terminal, a lightweight ore image classification algorithm based on cross-channel fine-grained feature fusion is proposed. A hybrid network is constructed by alternately using CNN and Transformer to effectively extract local and global information of the image. The cross-channel fine-grained feature fusion module is introduced as the feature fuser, and the fusion strategy of channel grouping and random channel shuffle is adopted to enhance the acquisition of ore texture information and maintain the diversity of fine-grained features. The multi-scale lightweight self-attention module is used to reduce the model parameters, enhance the perception of different scales and spatial locations, and ensure the stability of training and avoid overfitting low-level features. An efficient coordinate attention module is constructed as a fine-grained feature extractor to achieve lightweight and efficient feature extraction. The proposed algorithm achieves 95.78% and 94.77% classification accuracy on the two public ore image datasets of Mineral Photos and Petrology Thin Section Data on the Kaggle platform, respectively. Compared with the other nine lightweight classification networks, such as ShuffleNetV2, MobileNetV3, RegNet, ConvNeXtV2, LeViT, EdgeViTs, AFFNeT, EdgeNeXt and MViTV2. The proposed algorithm has fewer parameters (1.27?MB), lower computation (269?MFLOPs) and faster classification speed (219?FPS).

Key words: ore image classification, convolutional neural network (CNN), Transformer, cross-channel feature fusion, attention mechanism

高云霏, 吕伏, 冯永安. 跨通道细粒度特征融合的矿石图像分类算法[J]. 计算机工程与应用, 2025, 61(10): 214-227.

GAO Yunfei, LYU Fu, FENG Yong’an. Ore Image Classification Algorithm Based on Cross-Channel Fine-Grained Feature Fusion[J]. Computer Engineering and Applications, 2025, 61(10): 214-227.

参考文献

[1] 毕鹏程, 罗健欣, 陈卫卫. 轻量化卷积神经网络技术研究[J]. 计算机工程与应用, 2019, 55(16): 25-35.
BI P C, LUO J X, CHEN W W. Research on lightweight convolutional neural network technology[J]. Computer Engineering and Applications, 2019, 55(16): 25-35.
[2] 杜睿山, 陈雨欣, 孟令东, 等. 融合改进残差网络和注意力的黏土矿物图像分类[J]. 计算机工程与应用, 2024, 60(23): 333-339.
DU R S, CHEN Y X, MENG L D, et al. Clay mineral image classification based on improved residual network and attention[J]. Computer Engineering and Applications, 2024, 60(23): 333-339.
[3] 李文静, 白静, 彭斌, 等. 图卷积神经网络及其在图像识别领域的应用综述[J]. 计算机工程与应用, 2023, 59(22): 15-35.
LI W J, BAI J, PENG B, et al. Graph convolutional neural network and its application in image recognition[J]. Computer Engineering and Applications, 2023, 59(22): 15-35.
[4] ZHENG Q H, ZHAO P H, ZHANG D L, et al. MR-DCAE: manifold regularization-based deep convolutional autoencoder for unauthorized broadcasting identification[J]. International Journal of Intelligent Systems, 2021, 36(12): 7204-7238.
[5] 马金林, 张裕, 马自萍, 等. 轻量化神经网络卷积设计研究进展[J]. 计算机科学与探索, 2022, 16(3): 512-528.
MA J L, ZHANG Y, MA Z P, et al. Research progress of lightweight neural network convolution design[J]. Journal of Frontiers of Computer Science and Technology, 2022, 16(3): 512-528.
[6] LIU Y, WANG X Y, ZHANG Z L, et al. LOSN: lightweight ore sorting networks for edge device environment[J]. Engineering Applications of Artificial Intelligence, 2023, 123: 106191.
[7] 杨文龙, 郭明钰. 轻量级注意力X射线矿石检测方法[J]. 电子测量技术, 2022, 45(18): 71-79.
YANG W L, GUO M Y. Lightweight attention parallel X-ray ore detection algorithm[J]. Electronic Measurement Technology, 2022, 45(18): 71-79.
[8] SANDLER M, HOWARD A, ZHU M L, et al. MobileNetV2: inverted residuals and linear bottlenecks[C]//Proceedings of the 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2018: 4510-4520.
[9] ZHANG Y, CHENG L, PENG Y T, et al. Faster OreFSDet: a lightweight and effective few-shot object detector for ore images[J]. Pattern Recognition, 2023, 141: 109664.
[10] 程卫月, 张雪琴, 林克正, 等. 融合全局与局部特征的深度卷积神经网络算法[J]. 计算机科学与探索, 2022, 16(5): 1146-1154.
CHENG W Y, ZHANG X Q, LIN K Z, et al. Deep convolutional neural network algorithm fusing global and local features[J]. Journal of Frontiers of Computer Science and Technology, 2022, 16(5): 1146-1154.
[11] 彭斌, 白静, 李文静, 等. 面向图像分类的视觉Transformer研究进展[J]. 计算机科学与探索, 2024, 18(2): 320-344.
PENG B, BAI J, LI W J, et al. Survey on visual transformer for image classification[J]. Journal of Frontiers of Computer Science and Technology, 2024, 18(2): 320-344.
[12] 李建, 杜建强, 朱彦陈, 等. 基于Transformer的目标检测算法综述[J]. 计算机工程与应用, 2023, 59(10): 48-64.
LI J, DU J Q, ZHU Y C, et al. Survey of transformer-based object detection algorithms[J]. Computer Engineering and Applications, 2023, 59(10): 48-64.
[13] 徐光宪, 冯春, 马飞. 基于UNet的医学图像分割综述[J]. 计算机科学与探索, 2023, 17(8): 1776-1792.
XU G X, FENG C, MA F. Review of medical image segmentation based on UNet[J]. Journal of Frontiers of Computer Science and Technology, 2023, 17(8): 1776-1792.
[14] ZHENG Q H, ZHAO P H, WANG H J, et al. Fine-grained modulation classification using multi-scale radio transformer with dual-channel representation[J]. IEEE Communications Letters, 2022, 26(6): 1298-1302.
[15] TANG X X, WANG X L, YAN N, et al. A new ore image segmentation method based on Swin-Unet[C]//Proceedings of the 2022 China Automation Congress. Piscataway: IEEE, 2022: 1681-1686.
[16] 汤翔中, 高丙朋. 融合注意力空洞卷积和Transformer的矿石图像分割[J]. 科学技术与工程, 2023, 23(16): 6974-6982.
TANG X Z, GAO B P. Ore image segmentation based on attention hole convolution and transformer[J]. Science Technology and Engineering, 2023, 23(16): 6974-6982.
[17] HU J, SHEN L, ALBANIE S, et al. Squeeze-and-excitation networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020, 42(8): 2011-2023.
[18] HOWARD A, SANDLER M, CHU G, et al. Searching for MobileNetV3[J]. arXiv:1905.02244, 2019.
[19] GRAHAM B, EL-NOUBY A, TOUVRON H, et al. LeViT: a vision transformer in ConvNet’s clothing for faster inference[J]. arXiv:2104.01136, 2021.
[20] MA N N, ZHANG X Y, ZHENG H T, et al. ShuffleNet V2: practical guidelines for efficient CNN architecture design[J]. arXiv:1807.11164, 2018.
[21] MEHTA S, RASTEGARI M. Separable self-attention for mobile vision transformers[J]. arXiv:2206.02680, 2022.
[22] PENG Z L, HUANG W, GU S Z, et al. Conformer: local features coupling global representations for visual recognition[J]. arXiv:2105.03889, 2021.
[23] WADEKAR S N, CHAURASIA A. MobileViTv3: mobile-friendly vision transformer with simple and effective fusion of local, global and input features[J]. arXiv:2209.15159, 2022.
[24] CHEN J R, KAO S H, HE H, et al. Run, don’t walk: chasing higher FLOPS for faster neural networks[J]. arXiv:2303. 03667, 2023.
[25] HOWARD A G, ZHU M L, CHEN B, et al. MobileNets: efficient convolutional neural networks for mobile vision applications[J]. arXiv:1704.04861, 2017.
[26] ZHANG X Y, ZHOU X Y, LIN M X, et al. ShuffleNet: an extremely efficient convolutional neural network for mobile devices[J]. arXiv:1707.01083, 2017.
[27] LIU C X, ZOPH B, NEUMANN M, et al. Progressive neural architecture search[C]//Proceedings of the 15th European Conference on Computer Vision. Cham: Springer, 2018: 19-35.
[28] HE K M, ZHANG X Y, REN S Q, et al. Deep residual learning for image recognition[J]. arXiv:1512.03385, 2015.
[29] HUANG G, LIU S C, VAN DER MAATEN L, et al. CondenseNet: an efficient DenseNet using learned group convolutions[J]. arXiv:1711.09224, 2017.
[30] MAAZ M, SHAKER A, CHOLAKKAL H, et al. EdgeNeXt: efficiently amalgamated CNN-transformer architecture for mobile vision applications[J]. arXiv:2206.10589, 2022.
[31] DOSOVITSKIY A, BEYER L, KOLESNIKOV A, et al. An image is worth 16x16 words: transformers for image recognition at scale[J]. arXiv:2010.11929, 2020.
[32] GAO S H, CHENG M M, ZHAO K, et al. Res2Net: a new multi-scale backbone architecture[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021, 43(2): 652-662.
[33] HOU Q B, ZHOU D Q, FENG J S. Coordinate attention for efficient mobile network design[J]. arXiv:2103.02907, 2021.
[34] CHEN X N, LIANG C, HUANG D, et al. Symbolic discovery of optimization algorithms[J]. arXiv:2302.06675, 2023.
[35] RADOSAVOVIC I, KOSARAJU R P, GIRSHICK R, et al. Designing network design spaces[J]. arXiv:2003.13678, 2020.
[36] WOO S, DEBNATH S, HU R H, et al. ConvNeXt V2: co-designing and scaling ConvNets with masked autoencoders[J]. arXiv:2301.00808, 2023.
[37] PAN J T, BULAT A, TAN F W, et al. EdgeViTs: competing light-weight CNNs on mobile devices with vision transformers[J]. arXiv:2205.03436, 2022.
[38] HUANG Z P, ZHANG Z Z, LAN C L, et al. Adaptive frequency filters as efficient global token mixers[J]. arXiv:2307.14008, 2023.