基于PANet网络和先验知识的变压器表计图像识别算法

Transformer Meter Image Recognition Algorithm Based on PANet network and Prior Knowledge

本文中作者针对传统图像处理识别方法鲁棒性较差以及深度学习算法数据量需求大的问题,本文将先验知识引入PANet模型,提出了一种全新的电气表计识别技术。此外,为解决光照干扰问题,该框架还构建了基于感知理论的光照补偿模块。整体实现思路为:首先调整图像亮度,利用PANet网络检测表盘、指针、刻度等关键信息,再结合先验知识对上述信息进行修正,最终根据修正后的信息计算准确读数。实验表明,该方法在电力表计识别任务中相较于其它模型,具有更强的迁移性,且识别准确率明显提升,对电力设备图像识别这一小样本问题具有一定的借鉴意义。

To solve the problem of traditional image processing methods and deep learning algorithms,this paper proposes a new framework by combining prior knowledge and PANet model.In addition,to eliminate the effect of different illumination,an illumination compensation module based on perception theory is also added.The implementation is as follows:first adjusting the illumination of image,and then using PANet Model to detect key information such as positions of dial plate,needle and scales.The detection results will be modified based on prior knowledge,and finally be utilized to calculate the reading.Experiments show that the proposed method,compared with related works,is more robust and stable,and much higher in accuracy,which can be reference for the small-sample learning problem of power equipment recognition.