The development of modern high-altitude wide-swath imaging systems has brought about adaptive quantization bit-depths(10 ~ 16 bits) and large-scale datasets with a single frame approaching I0 GB,posing two major techn...The development of modern high-altitude wide-swath imaging systems has brought about adaptive quantization bit-depths(10 ~ 16 bits) and large-scale datasets with a single frame approaching I0 GB,posing two major technical challenges for thin-cloud removal in large-format aerial images.Firstly,it is difficult to construct a unified model across different bit-depths,resulting in poor model reusability and the need for high retraining costs in new domains.Secondly,traditional neural networks have to segment images into sub-blocks for processing and then splice them,which is prone to generating chromatic artifacts.To address these issues,we propose the Seamless Cloud Elimination Network(SCENet),whose core innovations are as follows:I achieving bit-depth unification through 8-bit standardization of paired images to support unified model training;2 adopting an adaptive transfer learning architecture that freezes encoder weights and fine-tunes decoders to realize efficient domain adaptation and rapid cloud removal;3 innovating a white-balance-aware cross-patch network architecture,which avoids chromatic artifacts during reconstruction while learning cloud features.Experiments show that this method performs excellently on real datasets,and SCENet achieves the highest Peak Signal-to-Noise Ratio(PSNR) compared with eight existing state-of-the-art methods.展开更多
文摘The development of modern high-altitude wide-swath imaging systems has brought about adaptive quantization bit-depths(10 ~ 16 bits) and large-scale datasets with a single frame approaching I0 GB,posing two major technical challenges for thin-cloud removal in large-format aerial images.Firstly,it is difficult to construct a unified model across different bit-depths,resulting in poor model reusability and the need for high retraining costs in new domains.Secondly,traditional neural networks have to segment images into sub-blocks for processing and then splice them,which is prone to generating chromatic artifacts.To address these issues,we propose the Seamless Cloud Elimination Network(SCENet),whose core innovations are as follows:I achieving bit-depth unification through 8-bit standardization of paired images to support unified model training;2 adopting an adaptive transfer learning architecture that freezes encoder weights and fine-tunes decoders to realize efficient domain adaptation and rapid cloud removal;3 innovating a white-balance-aware cross-patch network architecture,which avoids chromatic artifacts during reconstruction while learning cloud features.Experiments show that this method performs excellently on real datasets,and SCENet achieves the highest Peak Signal-to-Noise Ratio(PSNR) compared with eight existing state-of-the-art methods.
