Railway few-shot intruding objects detection method with metric meta learning

Table 1. Railway dataset composition
View table
View in Article
Table 1. Railway dataset composition
编号场景类别数量/张
1 空场景 40
2 行人入侵 40
3 泥石流入侵 40
4 落石入侵 40
5 列车经过 40

Table 1. [in Chinese]

View table

View in Article

Table 1. [in Chinese]

算法 1：类中心微调（Pseudo Code of center fine tuning）

输入：支持集 $S = {(f (x_{1}), y_{1}), \dots, (f (x_{N}), y_{N})}$ ，其中 $y_{i} \in \{1, . . ., k\}$ ， $D_{k}$ 表示支持集 $S$ 中所有标签为 $y_{i} = k$ 的样本集合

输出：更新得到的每个中心值 ${c_{1}, . . ., c_{k}}$

1.初始化每个类中心值：

for k in ${1, . . ., k}$ do

$c_{k} (φ) \leftarrow \frac{1}{N} \sum_{x_{i} \in S_{k}} f_{ϕ}^{k} (x_{i})$

end for

2.归一化支持集中样本x_i到每个初始类中心的距离：

for i in $\{1, . . ., N\}$ do

$p (φ) (y = k | x_{i}) \leftarrow \frac{e^{- d (f φ (x_{i}), c_{k} (φ))}}{\sum_{k^{'}} e^{- d (f_{φ} (x_{i}), c_{k}^{'} (φ))}}$

end for

3.计算损失并更新类中心：

$L (φ) \leftarrow - \frac{1}{N} \sum_{i = 1}^{N} l o g (p (φ) (y = k | x_{i}))$

Table 2. [in Chinese]

View table

View in Article

Table 2. [in Chinese]

算法2：特征映射网络学习算法

（Feature mapping network parameter learning algorithm）

输入：支持集 $S = {(f (x_{1}), y_{1}), \dots, (f (x_{N}), y_{N})}$ ，其中 $y_{i} \in \{1, \dots, k\}$ ，查询集 $Q = {(f (x_{1}), y_{1}), \dots,$ $(f (x_{m}), y_{m})}$ ，特征映射网络初始化参数 $θ_{C}$ ，超参数 $λ$ 和 $α$ ，迭代次数 $t \leftarrow 0$ 。

输出：特征映射网络参数 $θ_{C}$

1：while not converge do：

2： $t \leftarrow t + 1$

3：计算总的损失：

$L_{t o t a l} = L_{s} + L_{c} = - \frac{1}{m} \sum_{i = 1}^{m} l o g (\frac{e^{- d (f_{ϕ} (x_{i}), c_{k})}}{\sum_{k^{'}} e^{- d (f_{ϕ} (x), c_{k^{'}})}}) + λ \sum_{i = 1}^{m} {‖x_{i} - c_{k}‖}^{2}$

4：初始化类别中心：

for i in k：

$c_{k} = \frac{1}{|S_{k}|} \sum_{(x_{i}, y_{i}) \in S_{k}} f (x_{i})$

end for

5：计算反向传播误差： $\frac{\partial L}{\partial x_{i}} = \frac{\partial L_{s}}{\partial x_{i}} + λ \cdot \frac{\partial L_{c}}{\partial x_{i}}$

6：更新参数 $θ_{C}$ ： $θ_{C} = θ_{C} - α \sum_{i}^{m} \frac{\partial L}{\partial x_{i}} \cdot \frac{\partial x_{i}}{\partial θ_{C}}$

7：end while

Table 2. Experiment results on MiniImageNet dataset（%）
View table
View in Article
Table 2. Experiment results on MiniImageNet dataset（%）
5-way Acc
Model 1-shot 5-shot
MAM^［17］ 48.70 63.11
RelationNet^［9］
ProtoNet^［8］
MatchNet^［18］
本文算法
50.44
48.51±0.40
43.40±0.78
50.65±0.45
65.32
67.09±0.36
51.09±0.71
74.40±0.33

Table 3. Experiment results on railway datasets
View table
View in Article
Table 3. Experiment results on railway datasets
5-way Acc
Model 1-shot 5-shot
MAML 60.23±0.21 75.26±0.24
RelationNet
ProtoNet
MatchNet
本文算法
61.95±0.32
62.40±0.36
58.45±0.23
65.91±0.35
76.68±0.40
81.83±0.24
65.23±0.35
85.44±0.25

Table 4. mAp on different kinds of object
View table
View in Article
Table 4. mAp on different kinds of object
目标类别准确率
背景、泥石流、列车 95.53±0.18
背景、泥石流、列车、落石 91.48±0.21
背景、泥石流、列车、落石、行人 85.44±0.25

Table 5. Experiments on different attention mechanism（%）
View table
View in Article
Table 5. Experiments on different attention mechanism（%）
网络结构 5-way 1-shot 5-way 5-shot
基本网络 59.01±0.41 77.53±0.25
基本网络+CBAM 58.12±0.26 75.85±0.23
基本网络+改进CAM 62.76±0.41 78.39±0.24

Table 6. Ablation experiments on railway datasets

View table

View in Article

Table 6. Ablation experiments on railway datasets

基本网络	改进CAM	类中心微调	中心相关损失	模型预训练	5-way 1-shot	5-way 5-shot
√					59.01±0.41	77.53±0.25
√	√				62.76±0.41	78.39±0.24
√		√			-	80.58±0.24
√			√		64.90±0.39	79.44±0.22
√				√	64.16±0.40	84.23+0.23
√	√	√	√	√	65.91±0.35	85.44±0.25

Table 7. Model accuracy under different K-shot
View table
View in Article
Table 7. Model accuracy under different K-shot
N-way K-shot Accuracy
5-way 5-shot 85.44±0.25
5-way 10-shot 88.16±0.22
5-way15-shot 91.21±0.21
5-way 20-shot 93.51±0.21

Tools

Get Citation

Copy Citation Text

Baoqing GUO, Defen ZHANG. Railway few-shot intruding objects detection method with metric meta learning[J]. Optics and Precision Engineering, 2023, 31(12): 1816

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites