Abstract: An algorithm based on an improved YOLOv8n model was investigated to address the challenges posed by the low detection accuracy and high computational complexity of power equipment detection using infrared images in complex environments. First, the bottleneck structures in the backbone network C2F module were replaced with deformable convolutional network version 3 bottleneck modules to enhance the ability of the model to extract the features of power equipment with various shapes. Second, the coordinate attention (CA) mechanism was integrated to enhance the focus of the model on the infrared images of the power equipment and reduce background interference. Subsequently, a lightweight dilated feature pyramid network was introduced to replace the original neck network, which significantly reduced the parameters and computational complexity of the model while maintaining the recognition accuracy. Finally, the SCYLLA intersection over union (SIoU) loss function was introduced to improve the localization accuracy of the predicted bounding boxes by optimizing the aspect ratio penalty between the predicted and ground truth boxes. In an experiment, the proposed algorithm achieved a mean average precision of 94.92% on a power-equipment infrared image dataset, which was a 2.87% improvement over the baseline model. Moreover, the number of parameters and computational complexity were reduced by 43.3% and 24.7%, respectively. Compared with other mainstream detection algorithms, the proposed method exhibited significant advantages, thereby contributing meaningfully to the advancement of automated power-equipment detection.