Active learning for interactive satellite image change detection

Sahbi, Hichem; Deschamps, Sebastien; Stoian, Andrei

doi:10.48550/arxiv.2110.04250

“…Beside meta-learning, another approach to reduce the number of required training data points is active learning [33], [34], [35], [36], [37]. Active learning amounts to the process of choosing which samples should be annotated next and incrementally added to the training set [38].…”

Section: B Backgroundmentioning

confidence: 99%

“…In (35) real and imaginary parts of the transmitted symbol f IQ,τ (x τ [i]). Note that the constellation Xτ of the transmitted symbols xτ [i] is also composed of 16 points via (36).…”

Section: B Frequentist and Bayesian Meta-learning For Demodulationmentioning

confidence: 99%

Bayesian Active Meta-Learning for Reliable and Efficient AI-Based Demodulation

Cohen¹,

Park²,

Simeone³

et al. 2022

IEEE Trans. Signal Process.

View full text Add to dashboard Cite

Two of the main principles underlying the life cycle of an artificial intelligence (AI) module in communication networks are adaptation and monitoring. Adaptation refers to the need to adjust the operation of an AI module depending on the current conditions; while monitoring requires measures of the reliability of an AI module's decisions. Classical frequentist learning methods for the design of AI modules fall short on both counts of adaptation and monitoring, catering to one-off training and providing overconfident decisions. This paper proposes a solution to address both challenges by integrating meta-learning with Bayesian learning. As a specific use case, the problems of demodulation and equalization over a fading channel based on the availability of few pilots are studied. Meta-learning processes pilot information from multiple frames in order to extract useful shared properties of effective demodulators across frames. The resulting trained demodulators are demonstrated, via experiments, to offer better calibrated soft decisions, at the computational cost of running an ensemble of networks at run time. The capacity to quantify uncertainty in the model parameter space is further leveraged by extending Bayesian meta-learning to an active setting. In it, the designer can select in a sequential fashion channel conditions under which to generate data for meta-learning from a channel simulator. Bayesian active meta-learning is seen in experiments to significantly reduce the number of frames required to obtain efficient adaptation procedure for new frames.

show abstract

“…The method of modifying prediction results by introducing click weights provides an interactive idea for our change detection method based on deep learning. At present, the interaction process in the field of change detection is focused on the model training stage [11][12][13][14], that is, some features are extracted according to unsupervised classification, and then human-computer interaction is introduced in the training process to reduce the number of manually labeled samples. This kind of interactive strategy can only improve the accuracy of a model or reduce the number of manually labeled samples; however, it cannot modify the change detection results twice, and the interaction degree is limited, so it is difficult to achieve the expectation of fine detection results.…”

Section: Introductionmentioning

confidence: 99%

ICD: VHR-oriented interactive change detection algorithm

Jiang¹,

Zhou²,

Cao³

et al. 2023

Preprint

0

View full text Add to dashboard Cite

<p> In recent years, deep learning has become the mainstream development direction in the change detection field, and its accuracy and speed have also reached a high level. However, the change detection method based on deep learning cannot predict all the change areas accurately, and its application is limited due to local prediction defects. For this reason, we propose an interactive change detection network (ICD) for very high resolution (VHR) based on a deep convolution neural network. The network integrates positive and negative click information in the distance layer of the change detection network, and users can correct the prediction defects by adding clicks. We carried out experiments on the open source dataset WHU and LEVIR-CD. By adding clicks, their F1-scores can reach 0.920 and 0.912, respectively, which are 4.3% and 4.2% higher than the original network. To better evaluate the correction ability of clicks, we propose a set of evaluation index—click correction ranges, which is suitable for evaluating clicks, and we carry out experiments on the above models. The results show that the method of adding clicks can effectively correct the prediction defects and improve the result accuracy. </p>

show abstract

“…However, research on interactive change detection is still in its infancy. Most of the existing studies on interactive change detection focus on training semisupervised classification models to reduce the labeling of original samples [25][26][27][28]. This kind of interaction strategy can only improve the model accuracy or reduce the number of manually labeled samples, but it cannot modify the change detection results twice.…”

Section: Introductionmentioning

confidence: 99%

ICD: VHR-oriented interactive change detection algorithm

Jiang¹,

Zhou²,

Cao³

et al. 2023

Preprint

0

View full text Add to dashboard Cite

<p> In recent years, deep learning has become the mainstream development direction in the change detection field, and its accuracy and speed have also reached a high level. However, the change detection method based on deep learning cannot predict all the change areas accurately, and its application is limited due to local prediction defects. For this reason, we propose an interactive change detection network (ICD) for very high resolution (VHR) based on a deep convolution neural network. The network integrates positive and negative click information in the distance layer of the change detection network, and users can correct the prediction defects by adding clicks. We carried out experiments on the open source dataset WHU and LEVIR-CD. By adding clicks, their F1-scores can reach 0.920 and 0.912, respectively, which are 4.3% and 4.2% higher than the original network. To better evaluate the correction ability of clicks, we propose a set of evaluation index—click correction ranges, which is suitable for evaluating clicks, and we carry out experiments on the above models. The results show that the method of adding clicks can effectively correct the prediction defects and improve the result accuracy. </p>

show abstract

Active learning for interactive satellite image change detection

Cited by 4 publications

References 38 publications

Bayesian Active Meta-Learning for Reliable and Efficient AI-Based Demodulation

Bayesian Active Meta-Learning for Reliable and Efficient AI-Based Demodulation

ICD: VHR-oriented interactive change detection algorithm

ICD: VHR-oriented interactive change detection algorithm

Contact Info

Product

Resources

About