关键词: 深度学习, 少样本图像分类, 注意力机制, 全维动态卷积, 马氏距离, 推土机距离

Abstract:

The distribution propagation graph network（DPGN）is a few-shot image classification algorithm based on deep learning.Unfortunately，the DPGN algorithm completely ignores semantic information，which is important for fine-grained classification.Therefore，it delivers poor classification performances.This paper proposes a new Few-shot learning algorithm based on the DPGN algorithm，SinAM-FRN＿layer-ODConv-DM＆EMD＿Distribution Propagation Graph Network（SFOD＿DPGN）.

First，to address the inability to extract image features by the feature extraction module of the DPGN algorithm，the SimAM attention mechanism is integrated into four residual blocks of the feature extraction network ResNet12.The SimAM attention mechanism can generate three-dimensional weights for feature maps from both spatial and channel dimensions，and then aggregates the generated weights with the feature maps to enable the improved ResNet12 to learn more and richer image features；Second，in view that the normalization method of the ResNet12 is affected by the number of images selected in training，the combination of batch normalization and the ReLu activation function in the main path of each residual block of the ResNet12 is changed to the combination of the filter response normalization（FRN）and the threshold linear unit activation function（TLU）.Because of the FRN without mean operation，it easily leads to activation with arbitrary bias far from zero.If the FRN combines with the ReLu activation function，this bias has adverse effects on training.This paper employs the TLU after the FRN to address the problem.The SFOD＿DPGN algorithm improves the classification accuracy and ensures its inference speed.Then，it optimizes the classifier module of the DPGN algorithm.To solve poor classification performance of the classifier module，the full dimensional dynamic convolution（ODConv）is selected to replace the common convolution in the classifier module.The ODconv employs a linear combination of n convolutional kernels and parallel strategies to introduce multidimensional attention mechanisms for dynamic weighting，making the convolution operation dependent on the input.The ODconv improves the robustness of the SFOD＿DPGN algorithm.Finally，the DPGN algorithm uses the L2 distance measurement method in the classifier module，easily causing errors in calculating the distance between samples.Based on the characteristics of distance measurement methods，the Mahalanobis Distance（MD）is suitable for calculating the distance between samples（point graphs）.The Earth Moves’s Distance（EMD）distance ismore suitable for calculating the distance between distribution graphs.This paper uses the MD and EMD to replace the L2 in order to improve the ability of the classifier to measure the distance between samples.It improves the classification accuracy of the SFOD＿DPGN algorithm.

Experiments on the CUB-200-2011 dataset shows the SFOD＿DPGN algorithm is superior to the DPGN algorithm over 5way-1shot and 5way-5shot classification tasks.The accuracy improves by 7.97% and 2.66% respectively.Meanwhile，ablation experiments are performed for each part to verify the effect of the improved ResNet12 and the classifier module.Compared to the DPGN algorithm，after the SimAM attention mechanism is integrated into the ResNet12，the accuracy improves by 2.77% and 1.16% over 5way-1shot and 5way-5shot classification tasks respectively.Furthermore，after the improving the normalization method and activation function of the ResNet12，the accuracy is 5.00% and 2.04% higher respectively over 5way-1shot and 5way-5shot classification tasks.After the further replacement of the common convolution with the ODconv，the accuracy is up by 7.25% and 2.42% respectively over 5way-1shot and 5way-5shot classification tasks.Our experimental results demonstrate all improvements are effective to improve classification accuracy of the SFOD＿DPGN algorithm.