Robust and Efficient YOLOv8 with Multi-scale Attention for Infrared Image Recognition of Power Equipment

Infrared imaging technology has widespread applications in the operation and maintenance of substations. Timely acquisition of status information about power equipment can effectively prevent electrical faults caused by equipment overheating. Traditional infrared image anomaly detection primarily relies on manual analysis and judgment, which suffers from issues such as time-consuming processes and high workloads. Therefore, rapidly and accurately locating and identifying target equipment in infrared images holds significant research value for enhancing the intelligence level of substations. This study proposes an improved YOLOv8 object detection model and combines it with the MSRCR image enhancement algorithm for the identification of power equipment infrared images. At first, an electrical power equipment infrared image object detection dataset is constructed. The MSRCR method is used to enhance the original infrared images, addressing issues such as low contrast and pixel blurring in infrared images under rainy or foggy conditions, thereby improving the model's detection capability of power equipment under such weather conditions. Second, a novel multi-scale attention module is introduced into the YOLOv8 backbone network to extract multi-scale features from the input image using convolution kernels of different sizes, thereby enhancing the representational capability of the initial features. To further improve detection accuracy, the FocalLoss is adopted to address the classification challenges caused by the imbalance in infrared image data. Test results show that the average recognition accuracy for 8 types of power equipment reaches 96.31%, with a detection speed of 71 fps. The experimental results validate the effectiveness of the proposed method.