Block Multi-Dimensional Attention for Road Segmentation in Remote Sensing Imagery

Fractional derivative based weighted skip connections for satellite image road segmentation

“…It can be observed from the table that the BMDANet [71] had achieved the maximum values of mean IoU and F1-score as 0.8802 and 0.9363 respectively. The results clearly show that the proposed network achieved a higher mean IoU of 0.9062 as compared to BMDANet [71] and hence than the other mentioned segmentation methods. The proposed method improved the segmentation results of SOTA for ORD by 3% in terms of mean IoU.…”

“…The quantitative results of the final proposed network (α = 0.5) on ORD are compared with other state-of-the-art road segmentation networks viz. Deeplabv3 [34], MACUNet [63], DeeplabV3+ [34], D-LinkNet [39], UNet++ [45], and BMDANet [71] and the results are reported in the Table 5. The values of F1-score and mean IoU corresponding to the aforementioned methods are extracted from [71].…”

“…Deeplabv3 [34], MACUNet [63], DeeplabV3+ [34], D-LinkNet [39], UNet++ [45], and BMDANet [71] and the results are reported in the Table 5. The values of F1-score and mean IoU corresponding to the aforementioned methods are extracted from [71]. It can be observed from the table that the BMDANet [71] had achieved the maximum values of mean IoU and F1-score as 0.8802 and 0.9363 respectively.…”

“…The values of F1-score and mean IoU corresponding to the aforementioned methods are extracted from [71]. It can be observed from the table that the BMDANet [71] had achieved the maximum values of mean IoU and F1-score as 0.8802 and 0.9363 respectively. The results clearly show that the proposed network achieved a higher mean IoU of 0.9062 as compared to BMDANet [71] and hence than the other mentioned segmentation methods.…”

“…The images are of 0.21m spatial resolution per pixel (zoom level 19) with variable size, some being of size 4500 × 5500 while others being of 1000 × 2000. To compare the segmentation results with the SOTA [71], the split of database is set to be as 14 images for training, 3 images for validating, and 3 for testing the model. For training the proposed network, the 14 training images were randomly segmented into patches of size 512 × 512, then the obtained patches were augmented by applying five different operations: horizontal / vertical flipping and rotation at 90 • , 180 • and 270 • .…”