Author Affiliations
School of Information Engineering, Shenyang University of Chemical Technology, Shenyang, Liaoning 110020, Chinashow less
Fig. 1. Network architecture of improved U-Net
Fig. 2. Architecture of residual block
Fig. 3. Residual block with channel attention mechanism
Fig. 4. Structure of atrous spatial pyramid pooling
Fig. 5. Optical microscope with digital camera
Fig. 6. PCM images. (a) Original image; (b) labeled image
Fig. 7. Data augmentation results. (a) Training image; (b) image with blur effect; (c) image with rotation effect; (d) adjustment effect of image aspect ratio; (e) image with color space perturbation effect
Fig. 8. Curves of loss, accuracy, and intersection over union varying with iterations. (a)(d) Loss; (b)(e) accuracy; (c)(f) intersection over union
Fig. 9. Segmentation results of different algorithms. (a) Test images; (b) labeled images; (c) segmentation results of DeepLabv3+; (d) segmentation results of U-Net; (e) segmentation results of proposed algorithm
Method | Target | P | R | RIoU |
---|
U-Net | Floc | 0.8085 | 0.9720 | 0.7898 | Filamentous | 0.5115 | 0.7460 | 0.4355 | Mean | 0.6600 | 0.8590 | 0.6172 | DeepLabv3+ | Floc | 0.8767 | 0.9023 | 0.8007 | Filamentous | 0.5974 | 0.6628 | 0.4581 | Mean | 0.7371 | 0.7826 | 0.6296 | Our algorithm | Floc | 0.8836 | 0.9280 | 0.8269 | Filamentous | 0.6379 | 0.6850 | 0.4932 | Mean | 0.7608 | 0.8065 | 0.6601 |
|
Table 1. Evaluation indexes
Method | IoU |
---|
U-Net | 0.6172 | U-Net+ASPP | 0.6263 | U-Net+ResNet+attention | 0.6542 | U-Net+ResNet+attention+ASPP | 0.6601 |
|
Table 2. Influence of each module on whole performance