基于改进MMAL的细粒度图像分类研究  

Analysis of fine-grained image classification through improved MMAL

作　　者：李冰锋[1,2] 冀得魁 杨艺 Li Bingfeng;Ji Dekui;Yang Yi(School of Electrical Engineering and Automation,Henan University of Technology,Jiaozuo 454000,China;Henan Province Key Laboratory for Intelligent Detection and Control of Coal Mine Equipment,Jiaozuo 454003,China)

机构地区：[1]河南理工大学电气工程与自动化学院,焦作454000 [2]河南省煤矿装备智能检测与控制重点实验室,焦作454003

出　　处：《电子测量技术》2024年第17期172-179,共8页Electronic Measurement Technology

基　　金：河南省科技攻关项目(222102210230);河南理工大学博士基金(B2018-33)项目资助。

摘　　要：针对细粒度图像分类中目标区域难以精准定位及其内部细粒度特征难以识别的问题,提出了一种基于改进MMAL的细粒度图像分类方法。首先,利用形变卷积的感知区域可变性原理,动态地感知样本图像中不同尺度和形状的目标区域特征,从而增强网络对目标区域位置的感知能力。随后,采用GradCAM梯度回流的方法生成网络注意力热图,以减小特征背景噪声的干扰,实现对图像目标区域的精准定位。最后,提出位置感知空间注意力模块,通过融合坐标位置和双尺度空间信息,显著提升了网络对目标区域细粒度特征的提取能力。实验结果表明,与基线算法相比,该方法在CUB-200-2011、Stanford Car和FGVC-Aircraft三个公共数据集上分类精度分别提升了1.4%、1.5%、1.9%,该结果验证了所提方法的有效性。To address the challenges of accurately locating target regions and identifying fine-grained features in fine-grained image classification,we propose a fine-grained image classification method based on an improved multi-scale deformable convolution(MMAL).Firstly,by leveraging the variable receptive field principle of deformable convolution,our method dynamically adapts to different scales and shapes of target regions in sample images,enhancing the network′s ability to perceive the position of these regions.Subsequently,we utilize the Grad-CAM gradient backpropagation technique to generate network attention heatmaps,which reduces the interference from background noise and achieves precise localization of the image target regions.Finally,we introduce a position-aware spatial attention module that integrates coordinate positions and dual-scale spatial information,significantly improving the network′s capability to extract fine-grained features of the target regions.Experimental results demonstrate that,compared to baseline methods,our approach achieves improvements of 1.4%,1.5%,and 1.9%in classification accuracy on the CUB-200-2011,Stanford Car,and FGVC-Aircraft datasets,respectively,validating the effectiveness of the proposed method.

关 键 词：细粒度图像分类 多尺度形变分组 位置感知空间注意力 GradCAM热图定位 多分支 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] TN791[自动化与计算机技术—计算机科学与技术]