TY - GEN
T1 - Aerial Image Dehazing Network Compression
T2 - 2nd International Conference on Image, Vision and Intelligent Systems, ICIVIS 2022
AU - Liu, Lulu
AU - Meng, Zhijun
AU - Wang, Kaipeng
AU - Zhang, Jiahui
AU - Wang, Zichen
N1 - Publisher Copyright:
© 2023, The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd.
PY - 2023
Y1 - 2023
N2 - Dehazing Generative Adversarial Networks (Dehazing GANs) have been widely used in the field of haze-free images. However, due to the huge computational costs and memory usage of dehazing GANs, they are difficult to be applied on the edge devices with limited resources. Although there are many efforts to solve these issues, these compressed dehazing GANs can not meet the real-time requirements on edge devices. In this way, we propose an efficient multi-granularity distillation scheme for aerial image dehazing network compression (AIDNC). We are the first one to introduce online distillation to compress dehazing GANs, where the student model is compressed by lightweight residual structure and channel pruning. In AIDNC, the student generator removes the discriminator and is optimized by the guidance information provided by gradually promoted teacher generator. Experimental results on two dehazing datasets show that AIDNC achieves 17.43 × MACs and 218.41 × parameters compression with few loss for image quality. It is proved that AIDNC is an efficient framework for the aerial image real-time dehazing on edge devices, like micro-UAVs.
AB - Dehazing Generative Adversarial Networks (Dehazing GANs) have been widely used in the field of haze-free images. However, due to the huge computational costs and memory usage of dehazing GANs, they are difficult to be applied on the edge devices with limited resources. Although there are many efforts to solve these issues, these compressed dehazing GANs can not meet the real-time requirements on edge devices. In this way, we propose an efficient multi-granularity distillation scheme for aerial image dehazing network compression (AIDNC). We are the first one to introduce online distillation to compress dehazing GANs, where the student model is compressed by lightweight residual structure and channel pruning. In AIDNC, the student generator removes the discriminator and is optimized by the guidance information provided by gradually promoted teacher generator. Experimental results on two dehazing datasets show that AIDNC achieves 17.43 × MACs and 218.41 × parameters compression with few loss for image quality. It is proved that AIDNC is an efficient framework for the aerial image real-time dehazing on edge devices, like micro-UAVs.
KW - Depthwise separable convolution
KW - Generative adversarial network compression
KW - Image dehazing
KW - Knowledge distillation
UR - https://www.scopus.com/pages/publications/85152622397
U2 - 10.1007/978-981-99-0923-0_44
DO - 10.1007/978-981-99-0923-0_44
M3 - 会议稿件
AN - SCOPUS:85152622397
SN - 9789819909223
T3 - Lecture Notes in Electrical Engineering
SP - 442
EP - 453
BT - Proceedings of International Conference on Image, Vision and Intelligent Systems, ICIVIS 2022
A2 - You, Peng
A2 - Li, Heng
A2 - Chen, Zhenxiang
PB - Springer Science and Business Media Deutschland GmbH
Y2 - 15 August 2022 through 17 August 2022
ER -