With the continuous development of intelligent power systems and augmented reality (AR) technology, AR wearable devices demonstrate significant potential for enhancing efficiency in power maintenance. However, in actual maintenance scenarios, images transmitted from wearable devices are often challenged by factors like low brightness, blurriness, and cluttered backgrounds. But existing image recognition algorithms often face issues such as low recognition accuracy or inability to recognize such images. To tackle this issue, this paper presents an improved YOLOv8 object detection approach specifically designed for recognizing challenging images. The optimization include replacing Convolutional Neural Network (CNN) with Dynamic Convolutional Network (DCN) and introducing effective channel attention (ECA) and Convolutional Block Attention Module (CBAM). Experimental results demonstrate that the proposed optimization algorithm significantly improves the model’s ability to recognize challenging images. Compared to the original YOLOv8 algorithm, the improved YOLOv8 model demonstrates a 2.8% increase in accuracy on the test set employed in this research. The improved YOLOv8 algorithm also significantly enhances the model’s robustness and generalizability.