Chinese Journal of Liquid Crystals and Displays, Volume. 39, Issue 8, 1001(2024)
Remote sensing image land feature segmentation method based on lightweight DeepLabV3+
Figures & Tables(19)
Fig. 2. Comparison chart of receptive fields with different expansion rates
Fig. 10. Comparison chart of DeepLabV3+ ablation experiments（Legend of CCF Dataset）
Fig. 11. Comparison chart of DeepLabV3+ ablation experiments（Legend of Huawei Ascend Cup Dataset）
Table 1. Dataset details
View tableView in ArticleTable 1. Dataset details
数据集 分辨率/m 波段 标注类别 CCF2020 2 RGB 其他、耕地、林地、水域、建筑 华为昇腾杯 0.1~4 RGB 其他、水体、交通运输、建筑、植被、裸土
Table 2. MobileNetV2 network structure
View tableView in ArticleTable 2. MobileNetV2 network structure
Input Operator t c n s 2 242×3 conv2d - 32 1 2 1 122×32 bottleneck 1 16 1 1 1 122×16 bottleneck 6 24 2 2 562×24 bottleneck 6 32 3 2 282×32 bottleneck 6 64 4 2 142×64 bottleneck 6 96 3 1 142×96 bottleneck 6 160 3 2 72×160 bottleneck 6 320 1 1 72×320 conv2d 1×1 - 1 280 1 1 72×1 280 avgpool 7×7 - - 1 - 1×1×1 280 conv2d 1×1 - k -
Table 3. Training parameters
View tableView in ArticleTable 3. Training parameters
Keys Values Epoch 110 Batch_size 8 Init_lr 0.007 Min_lr 7e-5 optimizer_type sgd momentum 0.9 lr_decay_type cos
Table 4. Comparison of the number of parameters of different models
View tableView in ArticleTable 4. Comparison of the number of parameters of different models
方法 参数量/个 原始Xception 54 709 445 +MobileNetV2 5 814 037 +损失函数 5 874 037 +SE 5 854 997 引入最大池化 5 854 997 +ECA 5 855 000
Table 5. Experimental comparison of different attention modules on CCF dataset
View tableView in ArticleTable 5. Experimental comparison of different attention modules on CCF dataset
方法 IoU/% PA/% F1/% 速率/（it·s－1） 其他 耕地 林地 水域 建筑 mIoU 其他 耕地 林地 水域 建筑 mPA 基线 58.76 83.41 71.39 82.39 68.89 72.97 67.05 95.23 78.91 87.03 80.88 81.82 83.87 26.74 +CANet 87.08 83.43 72.77 83.19 68.31 73.09 66.49 94.50 79.62 90.12 81.43 82.43 83.89 25.88 +CBAMNet 59.78 84.04 71.79 83.11 67.52 73.25 70.42 94.94 79.63 88.09 75.01 81.62 84.05 25.74 +ECANet 59.08 84.18 72.01 82.81 68.25 73.27 68.77 95.31 78.87 88.15 80.35 82.29 84.12 26.45 +SENet 60.57 83.79 71.76 83.34 67.51 73.39 71.74 94.50 80.70 86.91 76.91 82.42 84.21 27.75
Table 6. Comparison of ablation experiments of improvement algorithm for DeeplabV3+ model on CCF dataset
View tableView in ArticleTable 6. Comparison of ablation experiments of improvement algorithm for DeeplabV3+ model on CCF dataset
方法 IoU/% Precision/% F1/% 速率/（it·s－1） 其他 耕地 林地 水域 建筑 mIoU 其他 耕地 林地 水域 建筑 Acc 原始Xception 50.53 82.34 72.79 81.60 63.88 70.23 59.30 91.11 88.53 91.49 71.07 86.36 80.43 13.42 +MobileNetV2 59.00 83.77 71.57 83.25 66.75 72.87 76.37 88.13 88.51 94.50 79.68 87.93 83.71 25.80 +损失函数 58.76 83.41 71.39 82.39 68.89 72.97 82.61 87.04 88.22 93.92 82.29 87.83 83.87 26.74 +SE 60.57 83.79 71.76 83.34 67.51 73.39 79.56 88.08 86.62 95.29 84.89 88.10 84.11 27.75 引入最大池化 60.60 84.15 72.06 82.78 67.45 73.41 82.68 87.98 88.20 93.78 79.11 88.26 84.23 27.78 +ECA 58.62 84.00 71.68 83.43 69.58 73.47 78.37 88.25 87.61 94.02 83.23 88.28 84.29 28.12
Table 7. Comparison of ablation experiment of improvement algorithm for DeeplabV3+ model on Huawei Ascend Cup dataset
View tableView in ArticleTable 7. Comparison of ablation experiment of improvement algorithm for DeeplabV3+ model on Huawei Ascend Cup dataset
方法 IoU/% Precision/% F1/% 速率/（it·s－1） 其他 水体 交通运输 建筑 植被 裸土 mIoU 其他 水体 交通运输 建筑 植被 裸土 Acc 原始Xception 37.70 64.46 77.00 52.27 23.95 34.53 48.32 46.26 80.26 90.00 65.33 33.63 46.18 78.64 63.58 2.99 +MobilNetV2 64.06 66.09 82.51 54.63 32.95 43.16 57.23 85.71 82.55 88.89 71.14 45.46 61.47 84.11 71.56 5.67 +损失函数 66.43 71.73 84.17 62.37 43.59 49.77 63.01 84.17 84.37 91.12 72.64 60.20 63.74 86.48 76.46 8.20 +SE 68.04 71.67 84.13 61.48 41.09 52.68 63.18 86.56 84.68 91.21 70.62 55.28 66.10 84.44 76.62 8.02 引入最大池化 68.71 71.39 83.94 62.64 41.58 50.80 63.28 85.82 84.14 90.77 74.70 57.49 66.71 86.42 76.88 8.07 +ECA 69.29 71.63 84.08 62.40 42.73 50.51 63.43 88.11 84.82 91.10 73.86 55.64 63.33 86.47 77.04 8.05
Table 8. Experimental comparison of different model on CCF dataset
View tableView in ArticleTable 8. Experimental comparison of different model on CCF dataset
方法 骨干
网络
IoU/% Precision/% F1/% 其他 耕地 林地 水域 建筑 mIoU 其他 耕地 林地 水域 建筑 Accuracy UNet VGG16 46.81 80.98 68.79 81.22 61.64 67.89 77.75 83.71 88.79 95.34 78.32 85.73 81.96 PSPNet ResNet50 45.31 80.42 69.95 78.17 61.41 67.05 75.72 84.79 87.85 88.98 78.14 85.25 80.58 MobileNetV2 54.80 83.64 72.53 78.92 65.56 71.11 71.35 89.78 89.61 87.18 73.52 87.23 82.67 HRNetV2 — 44.05 78.85 65.65 77.69 61.99 65.65 85.20 82.33 87.29 89.52 72.05 84.16 79.72 DL-Unet — 55.36 82.22 70.22 81.41 60.67 69.97 75.28 86.46 91.43 92.29 67.67 86.67 82.30 DeepLabV3+ Xception 50.53 82.34 72.79 81.60 63.88 70.23 76.37 88.13 88.51 94.50 79.68 87.93 82.32 MS-DeepLabV3+ ResNet50 56.07 82.82 71.21 82.73 65.06 71.58 74.37 86.88 90.91 93.72 75.84 87.24 83.20 本文方法 MobileNetV2 58.62 84.00 71.68 83.43 69.58 73.47 78.37 88.25 87.61 94.02 83.23 88.08 84.29
Tools
Get Citation
Copy Citation Text
Jing MA, Zhonghua GUO, Zhiqiang MA, Xiaoyan MA, Jialong LI. Remote sensing image land feature segmentation method based on lightweight DeepLabV3+[J]. Chinese Journal of Liquid Crystals and Displays, 2024, 39(8): 1001
Paper Information
Category: Image Segmentation
Received: Sep. 6, 2023
Accepted: --
Published Online: Sep. 27, 2024
The Author Email: Zhonghua GUO (guozhh@nxu.edu.cn)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20