基于注意力机制的红外图像和自然图像电力设备异常检测方法研究

Research on Infrared Image and Natural Image Power Equipment Anomaly Detection Based on Attention Mechanism

提出了一种基于注意力机制的红外图像和自然图像电力设备异常检测方法。该方法包括三个分支:自然图像、红外图像和注意力分支。前两个分支通过相似的网络提取各自的深层次特征,注意力分支用于强化关键的深度特征并且抑制不相关声特征。此外,还加入了权重共享和一致性损失,充分挖掘两种模态下关键的深度特征,弥补两种模态之间的差距。

It proposes a power equipment anomaly detection method based on combination of infrared image and natural image with attention mechanism. Specifically,the proposed model consists of three branches:a branch processing the natural image,a branch processing the infrared image and an attention module between the former branches. The first two branches extract their deep features of infrared image and natural image through similar networks respectively,and the attention module is used to boost the key deep features and suppress irrelevant noise features. In addition,we use sharing and consistency loss to fully utilize the key deep features of the two modes to make up for the gap between the two modes.