多尺度部分卷积混合空间注意力的高分遥感影像分类

Multi-scale partial convolution fusion hybrid spatial attention for high-resolution remote sensing classification

高分辨率遥感影像场景类别多且类间相似度高,针对已有的遥感分类模型在相似特征区分困难以及运行速度慢等问题,该文提出了一种基于多尺度与部分卷积的快速图像分类网络。首先,该方法通过不同核大小的普通卷积分别抽取不同尺度的目标特征,实现不同尺度特征的融合;其次,在空间注意力上添加带有权重的通道注意力分支,调整定位目标位置信息的通道权重,扩大了网络宽度,加强了定位重要目标显著性特征的能力;然后,将Inception网络中的3×3普通卷积替换为3×3部分卷积,构建多尺度部分卷积模块(MPCRes block),拓宽了ResNet block的宽度并且减小了模型参数量,提升了网络性能;最后,将MPCRes block与FasterNet block进行交叉堆叠,构建更深层次的网络,通过卷积与池化层不断进行下采样,直至将提取到的特征输入到Softmax分类器进行分类。在AID和VGoogle数据集上对该方法和其他基于深度学习的方法进行实验对比,结果表明,提出模型的参数量为18.61 M,FLOPs为2.58 G,在AID数据集两种不同比例训练样本下分类精度分别提高到97.35%和98.99%,在VGoogle数据集两种不同比例训练样本下分别提高到94.09%和97.21%,体现出该模型在遥感场景分类任务中更具有优势。

High-resolution remote sensing image scenes have many categories and high similarities between categories.There are difficulties in distinguishing similar features and improving running speed.A fast image classification network based on multi-scale and partial convolution is proposed in this paper.First,this method extracts target features of different scales through ordinary convolutions of different kernel sizes to achieve the fusion of features of different scales;second,adds a channel attention branch with weights to the spatial attention to adjust the location of the target position information.The channel weight expands the network width and strengthens the ability to locate the salient features of important targets;then,the 3×3 ordinary convolution in the Inception network is replaced with a 3×3 partial convolution to construct a multi-scale partial convolution module(MPCRes block),widens the width of the ResNet block and reduces the number of model parameters,improving network performance;finally,the MPCRes block and FasterNet block are cross-stacked to build a deeper network,which is continuously processed through convolution and pooling layers Downsampling until the extracted features are input to the Softmax classifier for classification.Experimental comparisons between this method and other deep learning-based methods were conducted on two data sets of AID and VGoogle with different training ratios.The results show that the parameter amount of the proposed model is 18.61 M,and the FLOPs are 2.58 G,in two different ratios in the AID data set The classification accuracy reached 97.35%and 98.99%respectively under the training samples,and reached 94.09%and 97.21%respectively under the two different proportions of training samples in the VGoogle data set.The overall effect is better than the comparison model,which reflects that the model is more effective in remote sensing scene classification tasks advantages.