Researching | FRKDNet：feature refine semantic segmentation network based on knowledge distillation

Chinese Journal of Liquid Crystals and Displays, Volume. 38, Issue 11, 1590(2023)

FRKDNet：feature refine semantic segmentation network based on knowledge distillation

Shi-yi JIANG¹, Yang XU^1,2、*, Dan-yang LI¹, and Run-ze FAN¹

Author Affiliations

¹College of Big Data and Information Engineering，Guizhou University，Guiyang 550025，China

²Guiyang Aluminum-magnesium Design and Research Institute Co.Ltd.，Guiyang 550009，China

show less

Abstract Get PDF(in Chinese)

Figures&Tables (13)

References (24)

Paper Information

Figures & Tables(13)

Fig. 1. Network structure of model

Download full size

View in Article

Fig. 2. Feature refine module

Download full size

View in Article

Fig. 3. Distance calculation module

Download full size

View in Article

Fig. 4. Results of hyperparameter ablation experiment

Download full size

View in Article

Fig. 5. Segmentation results of Cityscapes

Download full size

View in Article

Fig. 6. Comparison of segmentation accuracy of different categories

Download full size

View in Article

Fig. 7. Comparison of segmentation results of Pascal VOC validation

Download full size

View in Article

Fig. 8. Comparison of different backbone networks

Download full size

View in Article

Fig. 9. Segmentation results of Pascal VOC

Download full size

View in Article

Table 0. [in Chinese]

View in Article

Table 0. [in Chinese]

算法1：FRKDNet：基于知识蒸馏的特征提炼语义分割网络

算法，通过处理未经最终Softmax层的训练结果 $W_{T}$ 与 $W_{S}$ ，得到类间距离损失函数 $L_{i n t e r}$ 与类内距离损失函数 $L_{i n t r a}$ ，实现知识蒸馏。

输入：数据集 $D$ ，教师网络训练结果 $W_{T}$ ，学生网络初始结果 $W_{S}$

输出：类间距离损失函数 $L_{i n t e r}$ ，类内距离损失函数 $L_{i n t r a}$ ，交叉熵损失函数 $L_{c l s}$

1：while 训练循环 do

2：将 $W_{T}$ 、 $W_{S}$ 经过特征提炼模块得到 $W_{T}^{*}$ 、 $W_{S}^{*}$

3： $L_{i n t e r} \leftarrow 类间距离计算 (W_{T}^{*}, W_{S}^{*})$ —— $公式 (8)$

4： $L_{i n t r a} \leftarrow 类内距离计算 (W_{T}^{*}, W_{S}^{*})$ —— $公式 (9)$

5： $L \leftarrow$ $知识蒸馏损失函数$ $(L_{i n t e r}, L_{i n t r a}, L_{c l s})$ ——公式（10）

6：利用损失函数 $L$ 更新学生网络

7：end

Table 1. Ablation experiment
View table
View in Article
Table 1. Ablation experiment
骨干网络 $L_{c l s}$ 特征提炼模块 $L_{i n t e r}$ $L_{i n t r a}$ MIoU/%
T：ResNet101 77.74
S：ResNet18 √ 72.08
S：ResNet18 √ √ 74.33
S：ResNet18 √ √ √ 75.61
S：ResNet18 √ √ √ 75.02
S：ResNet18 √ √ √ √ 76.53

Table 2. Comparison of experimental results of different methods

View in Article

Table 2. Comparison of experimental results of different methods

方法	Params/M	FLOPs/G	MIoU/%
方法	Params/M	FLOPs/G	Pascal VOC	Cityscapes
T：ResNet101	56.0	1 870.0	77.67	77.74
S：ResNet18	14.07	270.5	72.15	72.08
+SKD^［16］			72.77（+0.62）	72.86（+0.78）
+SKDD^［17］			73.60（+1.45）	73.51（+1.43）
+IFVD^［9］			72.24（+0.09）	72.32（+0.24）
+CD^［10］			74.08（+1.93）	75.48（+3.4）
+Ours			74.19（+2.04）	76.53（+4.45）

Table 3. Comparative experiment of different methods in Cityscapes

View in Article

Table 3. Comparative experiment of different methods in Cityscapes

方法	Params/M	FLOPs/G	MIoU/%
ENet^［6］	0.358	3.612	58.3
ESPNet^［18］	0.364	4.422	60.3
ERFNet^［19］	2.067	25.60	68.0
ICNet^［7］	26.50	28.30	69.5
FCN^［20］	134.5	333.9	62.7
RefineNet^［21］	118.1	525.7	73.6
T：DeepLabV3+-ResNet101	56.0	1 870.0	77.74
S：DeepLabV3+-ResNet18	14.07	270.5	76.53

Tools

Get Citation

Copy Citation Text

Shi-yi JIANG, Yang XU, Dan-yang LI, Run-ze FAN. FRKDNet：feature refine semantic segmentation network based on knowledge distillation[J]. Chinese Journal of Liquid Crystals and Displays, 2023, 38(11): 1590

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Research Articles

Received: Jan. 10, 2023

Accepted: --

Published Online: Nov. 29, 2023

The Author Email: Yang XU (xuy@gzu.edu.cn)

DOI:10.37188/CJLCD.2023-0010

Topics