基于知识蒸馏的变电目标检测模型压缩及集成应用

Model Compression and Integration Applications of Substation Objects Detection Based on Knowledge Distillation

针对基于深度学习的智能识别模型在变电智能巡视装备本体集成应用时硬件资源受限的问题,本文提出了基于知识蒸馏的压缩与集成应用方法。该方法通过采用Detr模型对初始目标进行识别,再利用Deformable Detr算法对Detr模型进行压缩,使压缩率达到87.5%的同时,确保目标检测精度维持在较高水平,实现了目标检测模型在变电站巡检机器人本体上的有效集成应用。

Aiming at the problems of the existing intelligent recognition models based on deep learning,which have multiple parameters and are not suitable for deployment on substation intelligent patrol robots with limited hardware resources,a model compression and integration applications method based on knowledge distillation is proposed.Firstly,a Detr model is produced to detect substation objects.Secondly,Deformable Detr algorithm is used for compressing the original Detr model,which reduces the size of the model by 87.5%while keeping a comparatively high detection accuracy.Finally,the method is proven to be quite useful for integrating models compressed on substation intelligent patrol robots.