作者: Rongfang Wang , Fan Ding , Jia-Wei Chen , Licheng Jiao , Liang Wang
DOI: 10.1109/IGARSS39084.2020.9323964
关键词:
摘要: Recently, many convolution neural networks have been successfully employed in bitemporal SAR image change detection. However, most of those are too heavy where large memory necessary for storage and calculation. To reduce the computational spatial complexity facilitate detection on edge devices, this paper, we propose a lightweight network In proposed network, replace regular convolutional layers with bottlenecks, which will not increase number channels. Furthermore, employ dilated kernels few non-zero entries reduces FLOPs convlutional operators. Comparing traditional our be faster, less parameters. We verify two sets images. The experimental results show that can obtain comparable performance heavy-weight network.