Fig. 2
From: Pest detection in dynamic environments: an adaptive continual test-time domain adaptation strategy
The proposed CDPD. The image \({{\textbf {X}}}\) is subjected to augmentation by a strong augmenter before being classified, yielding \(\mathcal {L}_{bce}\). The total loss \(\mathcal {L}_{total}\) accounts for batch and domain differences. \(\mathcal {L}_{bce}\) is tasked with updating the classifier, while \(\mathcal {L}_{total}\) refines the augmenter to maximize image divergence without inducing class changes. A weak augmenter operates in parallel as well. During training, the main and auxiliary models exchange knowledge through the DAG module that follows each RepNCSPELAN4 layer. Due to the presence of multiple RepNCSPELAN4 layers in the model, with a DAG module situated post each layer, we omit the repetitive sections in the illustration for clarity. The total loss updates main model, auxiliary model and DAG module. In testing, the pseudo-label with the highest confidence from both models is selected for further model refinement