Fig. 1. Improved DeepLabv3+network
Fig. 2. Depthwise separable convolution
Fig. 3. Squeeze and excitation module
Fig. 4. Comparison of normalization methods. (a) BN; (b) GN
Fig. 5. Segmentation results of proposed algorithm. (a) Input images; (b) ground truth; (c) segmentation results
Parameter name | Parameter selection |
---|
Enter picture size | 769×769 | Loss function | Cross entropy | Optimizer | SGD | Batch size | 16 | Iteration | 18500 |
|
Table 1. Experimental parameters
Evaluation index | Xception | ResNet-50 | MobileNetv3 |
---|
mIoU /% | 78.18 | 79.20 | 72.94 | Validating time /ms | 113 | 234 | 15 | Parameter quantity /MB | 78.53 | 38.72 | 2.18 |
|
Table 2. Comparison of Deeplabv3+ model performance of different backbone networks
Convolution method in ASPP | Training time /h | mIoU /% |
---|
Standard convolution | 3.39 | 68.50 | Depthwise separable convolution in ASPP | 2.56 | 66.78 |
|
Table 3. Comparison of different convolution methods
Depthwise separable convolution in ASPP | SE | GN | mIoU /% |
---|
| | | 68.50 | √ | | | 66.78 | | √ | | 71.23 | √ | √ | | 70.17 | √ | √ | √ | 72.94 |
|
Table 4. Performance results for different modules on Cityscape
Depthwise separable convolution in ASPP | SE | GN | mIoU /% |
---|
| | | 54.06 | √ | | | 53.42 | | √ | | 57.71 | √ | √ | | 56.31 | √ | √ | √ | 58.76 |
|
Table 5. Performance results for different modules on Foggy Cityscape
Algorithm | Backbone network | mIoU /% | Parameter quantity /MB |
---|
PSPNet[14] | Paper source | 78.42 | 84.75 | SegNet[8] | Paper source | 57.95 | 29.46 | DeepLabv3+ | Xception | 78.18 | 78.53 | DeepLabv3+ | ResNet-50 | 79.20 | 38.72 | Proposed algorithm | MobileNetv3 | 72.94 | 2.18 |
|
Table 6. Comparison of performance of different algorithms
Algorithm | mIoU /% | Parameter quantity /MB |
---|
FSSNet[23] | 62.32 | 0.20 | Fast-SCNN[24] | 69.25 | 2.33 | ENet[25] | 60.43 | 0.37 | Proposed algorithm | 72.94 | 2.18 |
|
Table 7. Comparison of lightweight image semantic segmentation algorithms