Page 66 - IJEEE-2022-Vol18-ISSUE-1

62 | Hussein & Ali

were trained for 100 epochs with 500 patches of 72 images
where used 7 image in testing and 65 image used in training
for “Aerial Segmentation”.
Following hyper parameters are used number-channel=3 ,
number-classes= 6 ,optimizer is Adam , Learning rate is
001, Batch size is 17 .

TABLE I
Results Training the model (U-Net)

loss accuracy Val-loss Val-accuracy
0.54446 0.76 0.52749 0.77
0.52756 0.77 0.52701 0.77
0.53875 0.76 0.53456 0.76
0.52270 0.77 0.50296 0.78
0.55374 0.76 0.53489 0.77

TABLE II
Results Training the model (CNN)

loss accuracy Val-loss Val-accuracy
0.4892 0.9613 0.4533 0.9632
0.4539 0.9659 0.4502 0.9720
0.4397 0.9676 0.4302 0.9719
0.4307 0.9686 0.4302 0.9668
0.4227 0.9699 0.4556 0.9750

2) Evaluate the model

Fig. 9: Focal loss U-Net model Fig. 11: Cross-entropy on the left, focal loss in the middle,
and IoU loss on the right.
Fig.10: Training CNN Model
D. Testing the Model VI. CONCLUSIONS

Important to note that there were only 65 images for The proposed algorithm is U-Net, which has been applied
training and 7 for validation, so we can’t expect great results. to aerial images. It works to give different color to each
But this number of data is enough for our purpose. category, and it is possible to assign a category to each pixel
in the image, such as the label with the word car or plane, and
this is called semantic. The process of adding the
corresponding pixels is performed directly with the previous
operations, which are very smart operations, so it differs
from FCN. It has preserved its spatial information because it
contains the copy & crop process. We show that such a
network can be trained end-to-end from very few images and
outperforms the prior best method (fully convolutional
network) .Moreover, the network is fast. Segmentation of a
572x572 images taken less than a second on a recent GPU.
As a result, we obtained a matrix of the same dimensions for
the input image, so U-Net was applied to the aerial images,
and the process of prediction of pixels in the border region
was accurate and fast through the results that was applied by
the pytorch library.

CONFLICT OF INTEREST

The authors have no conflict of relevant interest to this
article.