Cascaded feature enhancement network model for real-time video monitoring of power system

Abstract

The application of real-time monitoring has been widely used to detect the safety and stability of the electric power system. Traditional monitoring relies heavily on human judgment and is impossible to detect status in real-time. Recently, with the development of deep learning, the object detection algorithm based on the deep convolutional neural network becomes a great option for realizing real-time monitoring applications of the power system. However, in power system scenarios, failed or unreal-time detection of abnormal conditions may cause a hazardous accident. To apply and optimize the object detection algorithm, issues such as multi-scale objects, class imbalance, and difficulty in balance speed and accuracy need to be addressed to improve the detection performance. Thus, we present a cascaded feature enhancement network model that combining attention mechanism, feature fusion scheme, and Cascaded Refinement Scheme. Attention mechanism and feature fusion scheme can help extract more effective feature information of multi-scale objects. Cascaded Refinement Scheme can effectively solve the problem of class imbalance. The whole model can well balanced in detect speed and accuracy. Experiments are performed on two benchmarks: PSA_Datasets and PASCAL VOC. Our method gets an absolute gain of 1.6% (300×300 input), 2.6% (512×512 input) in terms of mAP result of PSA_Datasets and 1% (300×300 input), 1.6% (512×512 input) in PASCAL VOC Dataset, compared to the best results of other SOTA detectors.