Research on Stable Cross-Domain Detection Method Based on Attention Mechanism

• Published in: 2023 28th International Conference on Automation and Computing (ICAC) pp. 1 - 6
• Main Authors: Chen, Junyufeng, Luo, Xiandong, Bai, Yan, Deng, Qingkai, Chen, Tianyu, Yang, Haotian
• Format: Conference Proceeding
• Language: English
• Published: IEEE 30.08.2023
• Subjects: attention mechanisms; Automation; Computational modeling; Convolutional neural networks; cross-domain; Feature extraction; Fluctuations; object detection; stability; Stability analysis; Training; YOLOv5
• DOI: 10.1109/ICAC57885.2023.10275300

When traditional convolutional neural network algorithms are applied to extract target features from a single environment in the source domain during cross-domain detection, these features are different from the main features of the object, causing undesirable effects when used for target domain detection. Combined with YOLOv5 and the attention module, this paper proposes a stable cross-domain detection method based on attention mechanism, which applies the attention mechanism to enhance the main features of the object and weaken the environment features to achieve the extraction of the main features of the object and improve the stability of the algorithm in cross-domain detection. Furthermore, this paper uses unmixed source domain information for training and two target domain data that are different from the source domain distribution for detection. Introduction of the SE attention module is demonstrated by comparison experiments to improve the accuracy by 39.54% compared to the classical Faster R-CNN. After normalization, the coefficient of variation is reduced by 83.52% compared to the classical Faster R-CNN and by 8.48% compared to YOLOv5s, indicating the improved detection stability. Experimental outcomes show that this method outperform the comparison method in cross-domain stability and are of strong application.