Ground-Based Cloud Image Segmentation Network Based on Improved MobileNetV2

Hongkun Bu; Shuai Chang; Ye Gu; Chunyu Guo; Chengbang Song; Wei Xu; Lü Tianyu; Wei Zhao; Shoufeng Tong

doi:10.3788/LOP240505

Laser & Optoelectronics Progress, Volume. 61, Issue 18, 1837011(2024)

Ground-Based Cloud Image Segmentation Network Based on Improved MobileNetV2

Hongkun Bu¹, Shuai Chang^1,3、*, Ye Gu¹, Chunyu Guo², Chengbang Song¹, Wei Xu³, Lü Tianyu³, Wei Zhao¹, and Shoufeng Tong¹

Author Affiliations

¹National Defense Key Discipline Laboratory on Space-Ground Laser Communication, Changchun University of Science and Technology, Changchun 130022, Jilin, China

²Henghui Optoelectronic Measurement Technology (Jilin) Co., Ltd., Changchun 130000, Jilin, China

³Changchun Institute of Optics, Fine Mechanics and Physics, Chinese Academy of Sciences, Changchun 130033, Jilin, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(12)

Fig. 1. CloudHS-Net model network structure

Download full size

Fig. 2. Inverse residual network structure of MobileNetV2

Download full size

Fig. 3. Receptive field of [2, 4, 8] dilation rate convolution kernel stack and receptive field of [1, 2, 6] dilation rate convolution kernel stack

Download full size

Fig. 4. The process of hybird dilated convolution splicing

Download full size

Fig. 5. Schematic diagram of the efficient channel attention network structure

Download full size

Fig. 6. Examples of SWIMSEG dataset and HHCL-Cloud dataset

Download full size

Fig. 7. Prediction results of different methods. (a) Input image; (b) mask; (c) K-means; (d) Otsu; (e) U-ResNet; (f) DeepLabV3+; (g) CloudSegNet; (h) MACNN; (i) proposed method

Download full size

Fig. 8. Comparison of all sky datasets. (a) Input; (b) mask; (c) U-ResNet; (d) DeepLabV3+; (e) CloudSegNet; (f) MACNN; (g) proposed method

Download full size

Table 1. Segmentation results of cloud images by different methods

View table

Table 1. Segmentation results of cloud images by different methods

Method	Precision	Accuracy	Recall	F1-score	Error rate	MIoU
K-means	90.53	83.19	77.25	83.36	16.81	71.48
Otsu	88.15	86.63	87.20	87.67	13.37	78.05
U-ResNet	86.73	86.87	86.91	86.79	13.13	76.68
CloudSegNet	91.42	90.91	90.44	90.75	9.09	83.09
DeepLabV3+	92.28	92.33	92.24	92.26	7.67	85.64
MACNN	92.70	92.60	92.56	92.63	7.40	86.28
CloudHS-Net	93.10	93.63	93.53	93.31	6.37	87.47

Table 2. Evaluation indexes of cloud image segmentation by different methods

View table

Table 2. Evaluation indexes of cloud image segmentation by different methods

Method	Precision	Accuracy	Recall	F1-score	Error rate	MIoU
U-ResNet	91.97	93.27	93.93	92.94	6.73	86.81
CloudSegNet	93.31	94.21	93.89	93.59	5.79	88.01
DeepLabV3+	92.57	93.46	92.90	92.74	6.54	86.54
MACNN	93.81	94.71	94.54	94.15	5.29	89.01
CloudHS-Net	94.74	95.51	95.57	95.15	4.49	89.86

Table 3. Total params, total GFLOPs, training time, and testing time of five kinds of networks
View table
Table 3. Total params, total GFLOPs, training time, and testing time of five kinds of networks
Method Total params /10⁶ Total GFLOPs /10⁹ Training time /h Testing time /s
U-ResNet 43.933 184.100 4.0 15
CloudSegNet 1.832 23.728 1.7 69
DeepLabV3+ 5.813 52.867 2.3 83
MACNN 56.8 94.5 3.7 114
CloudHS-Net 6.224 79.935 2.6 92

Table 4. ablation experiments of Hybrid Splitting and ECANet structures
View table
Table 4. ablation experiments of Hybrid Splitting and ECANet structures
Method Precision Accuracy Recall F1-score Error rate MIoU
MobileNetV2 91.70 91.66 91.46 91.56 8.34 84.45
+ ECA 92.70 93.55 92.99 92.84 6.45 86.72
+ HS 93.76 94.66 94.45 94.09 5.34 88.89
CloudHS-Net 94.74 95.51 95.57 95.15 4.49 89.86

Tools

Get Citation

Copy Citation Text

Hongkun Bu, Shuai Chang, Ye Gu, Chunyu Guo, Chengbang Song, Wei Xu, Lü Tianyu, Wei Zhao, Shoufeng Tong. Ground-Based Cloud Image Segmentation Network Based on Improved MobileNetV2[J]. Laser & Optoelectronics Progress, 2024, 61(18): 1837011

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites