Fusion of fractal geometric features Resnet remote sensing image building segmentation

Shengjun XU; Ruoxuan ZHANG; Yuebo MENG; Guanghui LIU; Jiuqiang HAN

doi:10.37188/OPE.20223016.2006

Optics and Precision Engineering, Volume. 30, Issue 16, 2006(2022)

Fusion of fractal geometric features Resnet remote sensing image building segmentation

Shengjun XU1...2,2, Ruoxuan ZHANG1,2,2,*, Yuebo MENG1,2,2, Guanghui LIU1,2,2 and Jiuqiang HAN1 |Show fewer author(s)

Author Affiliations

¹College of Information and Control Engineering，Xi 'an University of Architecture and Technology， Xi'an70055， China

²Faculty of Electronic and Information Engineering， Xi'an Jiaotong University， Xi'an710049， China

²Xi'an Key Labratory of Building Manufactaring Intelligent & Automation Technology， Xi'an710055， China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(16)

Fig. 1. Resnet101 network and residual module

Download full size

View in Article

Fig. 2. Overall structure of the proposed network

Download full size

View in Article

Fig. 3. Fractal dimension in atrous spatial pyramid pooling

Download full size

View in Article

Fig. 4. Fractal feature extraction process

Download full size

View in Article

Fig. 5. Deeply separable convolution attention fusion module

Download full size

View in Article

Fig. 6. Graph of loss decline during network training

Download full size

View in Article

Fig. 7. Comparison of local results for building extraction

Download full size

View in Article

Fig. 8. Local segmentation of building extraction under road interference conditions

Download full size

View in Article

Fig. 9. Building extraction local segmentation results under trees interference conditions

Download full size

View in Article

Fig. 10. Local segmentation results of buildings extracted under shadow interference

Download full size

View in Article

Table 1. Improved DBC algorithm steps

View table

View in Article

Table 1. Improved DBC algorithm steps

算法1 基于改进DBC的分形维数算法
输入：特征图 $F_{i n}$ ，将 $F_{i n}$ 中全体元素从1开始依次编号，设最大编号为 $s_{m a x}$ ，用 $f_{s}$ 表示 $F_{i n}$ 中第 $s$ 个元素， $G$ 为元素的灰度级；
输出：分形维数矩阵 $F_{f r a c t a l}$ ，用 $d_{s}$ 表示 $F_{f r a c t a l}$ 第 $s$ 个元素；
Step 1：利用padding填充矩阵 $F_{i n}$ 的边缘像素；
Step 2：For $s = 1$ to $s_{m a x}$ ，以 $f_{s}$ 为中心在 $F_{i n}$ 中进行区域划分，子区域大小设为 $W \times W$ ，计算该子区域分形维数 $d_{s}$ ；
Step 2.1：For $w = 2$ to $W / 2$ ，对该区域利用 $w \times w$ 的网格进行分割，计算分割尺度 $r = w / W$ ；
Step 2.1.1：对每个网格，将其在子区域内的位置记为 $(i, j)$ ，利用高度为 $h = w \times G / W$ 的盒子进行堆积；
Step 2.1.2：寻找网格内灰度的最大最小值 $g_{m a x}$ 、 $g_{m i n}$ ，计算网络差分盒维数 $n_{r} (i, j) = ⌈g_{m a x} / t'⌉ - ⌈g_{m i n} / t'⌉ + 1$ ；
Step 2.1.3：利用公式 $N_{r} = \sum_{i, j} n_{r} (i, j)$ 求子区域内所有网格的差分盒维数之和 $N_{r}$ ；
Step 2.2：用最小二乘法对多组 $l o g (N_{r})$ 和 $l o g (1 / r)$ 进行直线拟合，该直线斜率即为该点分形维数 $d_{s}$ ；
Step 3：若 $s = s_{m a x}$ ，输出遥感图像的分形维数矩阵 $F_{f r a c t a l}$ ，算法结束；否则返回Step 2继续迭代。

Table 2. Parameter settings and FD-ASPP output

View table

View in Article

Table 2. Parameter settings and FD-ASPP output

Conv	r	$σ$	FD-ASPP output	Output size
Conv1			-	128×128
Conv2	3，5，11，15	7，11，23，31	$F_{o u t}^{(1)}$	128×128
Conv3	3，5，11，15	7，11，23，31	$F_{o u t}^{(2)}$	64×64
Conv4	3，5，11	7，11，23	$F_{o u t}^{(3)}$	32×32
Conv5	3，5	7，11	$F_{o u t}^{(4)}$	16×16

Table 3. WHU Building Dataset performance comparison

View table

View in Article

Table 3. WHU Building Dataset performance comparison

Model	Precision	Recall	F1-score	mIoU	Params/M	FLOPs/G
FCN^［6］	90.56%	90.40%	90.48%	88.87%	15.3	80.51
Segnet^［7］	92.05%	91.43%	91.74%	90.21%	29.4	160.56
Deeplab V3^［8］	92.83%	93.07%	92.95%	91.89%	15.3	64.59
U-net^［9］	93.69%	93.14%	93.41%	92.57%	17.2	160.33
SETR^［10］	91.95%	92.22%	92.08%	91.92%	63.9	34.76
AlignSeg^［11］	94.25%	94.56%	94.27%	94.19%	66.8	123.84
Our model	94.48%	94.62%	94.55%	94.15%	204	95.56

Table 4. Influence of different layers of FD-ASPP on segmentation index

View table

View in Article

Table 4. Influence of different layers of FD-ASPP on segmentation index

Resnet101	Layer1	Layer2	Layer3	Layer4	Precision	Recall	mIoU
√					91.41%	92.43%	91.27%
√	√				91.89%	91.68%	91.69%
√		√			92.56%	93.63%	92.23%
√			√		92.23%	93.74%	91.88%
√				√	92.81%	93.45%	92.84%
√	√	√	√	√	93.26%	94.03%	93.80%

Table 5. Ablation of void convolutional modules in the WHU Building Dataset
View table
View in Article
Table 5. Ablation of void convolutional modules in the WHU Building Dataset
No. Base-line FD-ASPP DS-CAF Precision Recall mIoU
1 √ × × 91.41% 92.43% 91.27%
2 √ √ × 93.26% 94.03% 93.80%
3 √ × √ 92.89% 93.56% 93.22%
4 √ √ √ 94.48% 94.62% 94.15%

Table 6. Comparative experimental results in different scenarios

View table

View in Article

Table 6. Comparative experimental results in different scenarios

干扰类型	Model	Precision	Recall	mIoU
道路干扰	FCN^［6］	88.09%	87.72%	85.97%
	Segnet^［7］	89.44%	89.25%	88.53%
	DeeplabV3^［8］	90.61%	90.30%	89.37%
	U-net^［9］	92.43%	92.64%	92.15%
	SETR^［10］	90.85%	91.12%	90.43%
	AlignSeg^［11］	93.96%	94.15%	93.26%
	Our model	94.27%	94.43%	93.99%
树木干扰	FCN^［6］	85.85%	86.27%	85.73%
	Segnet^［7］	87.72%	86.90%	86.34%
	DeeplabV3^［8］	89.54%	89.16%	88.42%
	U-net^［9］	91.14%	91.89%	91.33%
	SETR^［10］	88.29%	88.60%	87.38%
	AlignSeg^［11］	91.44%	92.05%	91.74%
	Our model	93.07%	92.41%	92.11%
建筑物阴影干扰	FCN^［6］	87.78%	87.02%	85.76%
	Segnet^［7］	89.19%	89.34%	88.55%
	DeeplabV3^［8］	89.01%	89.47%	88.91%
	U-net^［9］	91.90%	92.38%	91.29%
	SETR^［10］	89.43%	90.66%	90.02%
	AlignSeg^［11］	93.43%	94.20%	93.25%
	Our model	93.50%	94.06%	93.19%

Tools

Get Citation

Copy Citation Text

Shengjun XU, Ruoxuan ZHANG, Yuebo MENG, Guanghui LIU, Jiuqiang HAN. Fusion of fractal geometric features Resnet remote sensing image building segmentation[J]. Optics and Precision Engineering, 2022, 30(16): 2006

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites