Multi-Modal Cross-Attention Learning on Detecting Anomalies of Electrical Equipment

Currently, in field of electrical power system, techniques of anomalies detection are constantly innovating. Applications of neural network on processing of patrol images spare analyzers plenty of time, but subject to the relatively low resolution of the object contour under heavy weather, results do not show well on recognition of anomalies in electrical equipment, while multi-modal methods can import more information to the objects detected, thus may improve the success rate of capture. In this paper, we propose a feature-fusion model which uses cross-attention learning method to augment features of the anomalies with text of corresponding description and environment condition. After comparing experiments on self-constructed datasets of images and text, our model has achieved the state of art on multiple metrics. More importantly, it is found that adding additional features to the model can achieve better results through ablation experiments, which shows our model is scalable for a better solution.