TY - GEN
T1 - Change Detection from Areal Imagery Drones Using Siamese U-Net with Spatial Attention Module
AU - Khalid, Lalu Syamsul
AU - Jati, Grafika
AU - Caesarendra, Wahyu
AU - Jatmiko, Wisnu
N1 - Funding Information:
ACKNOWLEDGMENT This work was supported by Hibah Penelitian Dasar Kompetitif Nasional (PDKN) from Ministry of Education, Culture, Research, and Technology entitled “Pengembangan Sistem Pertahanan Udara dari Serbuan Drone Menggunakan Koordinasi Drone Berkelompok Berbasis Kecerdasan Artifisial”,No. NKB-904/UN2.RST/HKP.05.00/2022.
Publisher Copyright:
© 2022 IEEE.
PY - 2022
Y1 - 2022
N2 - This research discusses the development of a new model for task change detection. Siamese Neural networks with U-Net as basic architecture are combined with spatial attention modules to perform task change detection. This model is developed to get a lightweight model with good performance. In the implementation, there is no need to use enormous resources. To benchmark the model, we used the LEVIR-CD dataset, where this dataset has two paired images taken at different times. The information contained in the two paired images is that there are changes such as the presence of buildings such as houses that increase or decrease in a certain area during the time of taking the two images. We compared the proposed model with U-Net and Siamese U-Net without spatial attention modules to see how they differ in performance. Then, We also compared the F1 Score with the baseline model of the LEVIR-CD dataset. After hyperparameter tuning with epochs of 100 is performed, the result is that the F1 Scores tested can balance the baseline model with a faster training time.
AB - This research discusses the development of a new model for task change detection. Siamese Neural networks with U-Net as basic architecture are combined with spatial attention modules to perform task change detection. This model is developed to get a lightweight model with good performance. In the implementation, there is no need to use enormous resources. To benchmark the model, we used the LEVIR-CD dataset, where this dataset has two paired images taken at different times. The information contained in the two paired images is that there are changes such as the presence of buildings such as houses that increase or decrease in a certain area during the time of taking the two images. We compared the proposed model with U-Net and Siamese U-Net without spatial attention modules to see how they differ in performance. Then, We also compared the F1 Score with the baseline model of the LEVIR-CD dataset. After hyperparameter tuning with epochs of 100 is performed, the result is that the F1 Scores tested can balance the baseline model with a faster training time.
KW - Change Feature
KW - Siamese Neural Networks
KW - Spatial Attention
KW - U-Net
UR - http://www.scopus.com/inward/record.url?scp=85141833656&partnerID=8YFLogxK
U2 - 10.1109/IWBIS56557.2022.9924971
DO - 10.1109/IWBIS56557.2022.9924971
M3 - Conference contribution
AN - SCOPUS:85141833656
T3 - IWBIS 2022 - 7th International Workshop on Big Data and Information Security, Proceedings
SP - 51
EP - 58
BT - IWBIS 2022 - 7th International Workshop on Big Data and Information Security, Proceedings
PB - Institute of Electrical and Electronics Engineers Inc.
T2 - 7th International Workshop on Big Data and Information Security, IWBIS 2022
Y2 - 1 October 2022 through 3 October 2022
ER -