Method, device, and storage medium for deep learning based domain adaptation with data fusion for aerial image data analysis
Embodiments of the present disclosure provide a method, a device, and a storage medium for domain adaptation for efficient learning fusion (DAELF). The method includes acquiring data from a plurality of data sources of a plurality of sensors; for each of the plurality of sensors, training an auxiliary classifier generative adversarial network (AC-GAN) by a hardware processor with data from each data source of the plurality of data sources, thereby obtaining a trained feature extraction network and a trained label prediction network for each data source; forming a decision-level fusion network or a feature-level fusion network; and training the decision-level fusion network or the feature-level fusion network with a source-only mode or a generate to adapt (GTA) mode; and applying the trained decision-level fusion network or the trained feature-level fusion network to detect a target of interest.
1. A domain adaptation for efficient learning fusion (DAELF) method, comprising:
acquiring data from a plurality of data sources of a plurality of sensors;
for each of the plurality of sensors, training an auxiliary classifier generative adversarial network (AC-GAN) by a hardware processor, wherein the AC-GAN includes a feature extraction network, a label prediction network, a generator network, and a discriminator network, with data from each data source of the plurality of data sources, thereby obtaining a trained feature extraction network and a trained label prediction network for each data source;
using the trained feature extraction network and the trained label prediction network for each data source on a sensor side, and a corresponding centralized fusion network on a fusion center side to form a decision-level fusion network; or using the trained feature extraction network for each data source on the sensor side and a corresponding centralized fusion network on the fusion center side to form a feature-level fusion network;
training the decision-level fusion network or the feature-level fusion network with a source-only mode or a generate to adapt (GTA) mode, wherein:
at the source-only mode, the trained feature extraction network for each data source and the corresponding centralized fusion network are trained with labeled source data, and
at the GTA mode, the trained feature extraction network for each data source and the corresponding centralized fusion network are trained separately, wherein the trained feature extraction network for each data source is trained with the labeled source data and unlabeled target data; and the corresponding centralized fusion network is trained with the labeled source data only; and
applying the trained decision-level fusion network or the trained feature-level fusion network to detect a target of interest.
2. The method according to claim 1 , wherein training the AC-GAN includes:
inputting a source sample of the data from each data source into the feature extraction network for each data source to generate an embedding feature used by both the label prediction network and the generator network for each data source; and
inputting a target sample of the data from each data source into the feature extract network for each data source to generate an embedding feature only used by the generator network for each data source.
3. The method according to claim 1 , wherein:
at a training phase, for each data source, the AC-GAN has a stream 1 , including the feature extraction network and the label prediction network, and a stream 2 , including the feature extraction network, the generator network, and the discriminator network.
4. The method according to claim 1 , further including:
displaying the target of interest detected by the trained decision-level fusion network or the trained feature-level fusion network.
5. A domain adaptation for efficient learning fusion (DAELF) device, comprising:
a memory, configured to store program instructions for performing a DAELF method; and
a processor, coupled with the memory and, when executing the program instructions, configured for:
acquiring data from a plurality of data sources of a plurality of sensors;
for each of the plurality of sensors, training an auxiliary classifier generative adversarial network (AC-GAN) by a hardware processor, wherein the AC-GAN includes a feature extraction network, a label prediction network, a generator network, and a discriminator network, with data from each data source of the plurality of data sources, thereby obtaining a trained feature extraction network and a trained label prediction network for each data source;
using the trained feature extraction network and the trained label prediction network for each data source on a sensor side, and a corresponding centralized fusion network on a fusion center side to form a decision-level fusion network; or
using the trained feature extraction network for each data source on the sensor side and a corresponding centralized fusion network on the fusion center side to form a feature-level fusion network;
training the decision-level fusion network or the feature-level fusion network with a source-only mode or a generate to adapt (GTA) mode, wherein:
at the source-only mode, the trained feature extraction network for each data source and the corresponding centralized fusion network are trained with labeled source data, and
at the GTA mode, the trained feature extraction network for each data source and the corresponding centralized fusion network are trained separately, wherein the trained feature extraction network for each data source is trained with the labeled source data and unlabeled target data; and the corresponding centralized fusion network is trained with the labeled source data only; and
applying the trained decision-level fusion network or the trained feature-level fusion network to detect a target of interest.
6. The device according to claim 5 , wherein training the AC-GAN includes:
inputting a source sample of the data from each data source into the feature extraction network for each data source to generate an embedding feature used by both the label prediction network and the generator network for each data source; and
inputting a target sample of the data from each data source into the feature extract network for each data source to generate an embedding feature only used by the generator network for each data source.
7. The device according to claim 5 , wherein:
at a training phase, for each data source, the AC-GAN has a stream 1 , including the feature extraction network and the label prediction network, and a stream 2 , including the feature extraction network, the generator network, and the discriminator network.
8. The device according to claim 5 , wherein the method further includes:
displaying the target of interest detected by the trained decision-level fusion network or the trained feature-level fusion network.
9. A non-transitory computer-readable storage medium, containing program instructions for, when being executed by a processor, performing a domain adaptation for efficient learning fusion (DAELF) method, the method comprising:
acquiring data from a plurality of data sources of a plurality of sensors;
for each of the plurality of sensors, training an auxiliary classifier generative adversarial network (AC-GAN) by a hardware processor, wherein the AC-GAN includes a feature extraction network, a label prediction network, a generator network, and a discriminator network, with data from each data source of the plurality of data sources, thereby obtaining a trained feature extraction network and a trained label prediction network for each data source;
using the trained feature extraction network and the trained label prediction network for each data source on a sensor side, and a corresponding centralized fusion network on a fusion center side to form a decision-level fusion network; or using the trained feature extraction network for each data source on the sensor side and a corresponding centralized fusion network on the fusion center side to form a feature-level fusion network;
training the decision-level fusion network or the feature-level fusion network with a source-only mode or a generate to adapt (GTA) mode, wherein:
at the source-only mode, the trained feature extraction network for each data source and the corresponding centralized fusion network are trained with labeled source data, and
at the GTA mode, the trained feature extraction network for each data source and the corresponding centralized fusion network are trained separately, wherein the trained feature extraction network for each data source is trained with the labeled source data and unlabeled target data; and the corresponding centralized fusion network is trained with the labeled source data only; and
applying the trained decision-level fusion network or the trained feature-level fusion network to detect a target of interest.
10. The storage medium according to claim 9 , wherein training the AC-GAN includes:
inputting a source sample of the data from each data source into the feature extraction network for each data source to generate an embedding feature used by both the label prediction network and the generator network for each data source; and
inputting a target sample of the data from each data source into the feature extract network for each data source to generate an embedding feature only used by the generator network for each data source.
11. The storage medium according to claim 9 , wherein:
at a training phase, for each data source, the AC-GAN has a stream 1 , including the feature extraction network and the label prediction network, and a stream 2 , including the feature extraction network, the generator network, and the discriminator network.
12. The storage medium according to claim 9 , wherein the method further includes:
displaying the target of interest detected by the trained decision-level fusion network or the trained feature-level fusion network.