Failure Detection Of Infrared Thermal Imaging Power Equipment Based On Improved DenseNet

Timely maintenance of the power equipment is the key to ensure the normal operation of the transmission equipment. In the substation scenario, a thermal infrared image detection method is proposed for target detection to detect potential faults in advance. The proposed method replaces the backbone network of Faster RCNN with DenseNet, to extract richer features, and in order to reduce the number of parameters of the backbone network, replaces standard convolution with learnable group convolution. To alleviate the problem of feature loss in packet convolution, the SFR structure is added to activate the features and improve the feature utilization. In order to reduce the complexity of the network and reasonably reduce the number of convolutions, we obtain a better lightweight model, and improve the NMS algorithm for the problem of regional overlap detection omission. Experiments show that the algorithm used has higher accuracy than YOLO and SSD, and the improved model not only reduces the network complexity, but also improves certain performance, and the final detection accuracy is 95.8%, which can be well applied to thermal infrared image detection