Lightweight Neural Network-Based Infrared Image and Anomalous Heat Region Recognition for Electrical Equipment

In order to accurately and quickly recognize electrical devices and abnormal heating parts under infrared conditions, we propose a lightweight neural network-based method for identifying electrical devices and their abnormal heating parts. First, we construct a fast target detection network for real-time detection. Second, we propose an efficient feature retention module to avoid the feature loss problem in multi-layer convolutional operations. In addition, we introduce an attention mechanism based on the coordinate information interaction strategy to localize the target more accurately. Finally, we use a lightweight weighted bidirectional feature pyramid network to improve the detection accuracy at different scales. Experimental results on a dataset of 12109 electrical infrared images containing six types of electrical devices and anomalously heated areas show that the proposed method achieves good results, with an average recognition accuracy of 93.9% and a processing speed of 222 frames/second.