Few-Shot Industrial Meter Detection Based on Sim-to-Real Domain Adaptation and Category Augmentation

Rapid and accurate detection of industrial meters in complex scenarios is an essential step toward inspection robot automatic meter recognition. Deep learning (DL) is a promising solution. However, due to the lack of large-scale public industrial meter image datasets, it is very difficult to train industrial meter detection models based on DL. Therefore, in this article, we combine the image generation technique and sim-to-real domain adaption technique to address the problem of few-shot industrial meter detection in complex scenarios. Specifically, we use Stable Diffusion to generate abundant virtual samples as the source domain dataset by inputting textual prompts. A small number of real samples are used as the target domain dataset. In addition, to attenuate the effect of domain shift, we propose a domain adaptation object detection framework based on category augmentation. This framework introduces domain information into the classifier and combines uncertainty estimation, which not only eliminates the training of domain classifiers in traditional adversarial learning-based domain adaptation algorithms but also facilitates feature alignment between source domain and target domain. Experiments show that the framework achieves 50.8% mAP50:95 and $55.0\% F1$ score, which outperforms the network trained with only real images by 8.3% mAP50:95 and $8.7\% F1$ score. We can achieve close performance with only 25% of the target domain samples with the help of the source domain dataset. Moreover, our method also outperforms other state-of-the-art methods in supervised domain adaptation object detection.