Twin-Interaction Network: A Classification Method Based on Open Set Recognition and Periodic Full Retraining

In the industrial field, traditional classification models are unable to recognize unknown targets in real-time and lack the capability to accept and learn new knowledge in real-time. To address this isFsue, we propose a detection method based on Open Set Recognition and a dual-branch switching mechanism for periodic full retraining, called Twin-Interaction Network. By adding a shallow model on top of the Vision Transformer network for target feature extraction, this approach performs spatial transformations on feature vectors to enhance the accuracy of unknown target detection. Furthermore, to endow the model with the capability of periodic full retraining, two open-set recognition modules are integrated into a twin architecture, simultaneously performing classification and learning tasks without interference. The open-set recognition method is incorporated into the offline full-model retraining framework for validation. Through an alternating learning mechanism, the Twin-Interaction Network can acquire more precise feature tensors of unknown category, thereby making the classification accuracy of the model for unknown categories consistent with that for known categories. The generalization performance of the proposed method is evaluated using multiple public datasets under standard dataset settings, and the experimental results demonstrate that our method outperforms existing approaches in terms of open-set recognition accuracy.