可变形特征融合网络的设计及在复杂天气电力设备图像处理中的应用

Design of Deformable Feature Fusion Network and Its Application in Image Processing of Complex Weather Power Equipment

针对复杂天气环境下电力设备图像模糊而造成电力设备识别与检测受到严重影响的问题,设计了可变形特征融合网络来实现图像清晰化,从而提高复杂天气环境下电力设备的检测精度。将可变形注意力作为编码器-解码器组件引入U-Net架构中。将特征融合中间件作为连接上下文信息的桥梁,以此整合上下文信息并保留更多的空间细节信息。以指针式仪表图像为例,进行了指针式仪表图像去雨和指针式仪表自动读数实验。实验结果表明,所设计网络在指针表图像去雨方面有良好的效果,并提高了去雨后指针表的读数精度。

To solve the problem that power equipment recognition and detection are seriously affected by fuzzy images in complex weather environment,a deformable feature fusion network is designed to achieve image sharpness,so as to improve the detection accuracy of power equipment in complex weather environment.The deformable attention is introduced into U-Net architecture as an encoder-decoder component.Feature fusion middleware is used as a bridge to connect context information,so as to integrate context information and retain more spatial details.Taking pointer instrument image as an example,the experiments of image deraining and automatic reading of pointer instruments are carried out.The experimental results show that the designed network has a good effect on the rain removal of the pointer instrument image and improves the reading accuracy of the pointer instrument after the rain removal.