Automatically detecting insulator defects is a challenging task, especially against complex backgrounds that often lead to low recognition rates. To tackle this problem, we propose GAC-YOLO, an improved detection method built upon the YOLOv7 framework. Our approach begins by optimizing the GSConv module, which effectively reduces computational costs and memory usage without compromising accuracy. We then integrate the ACmix attention mechanism, combining the strengths of both convolution and self-attention to better focus on critical features. Furthermore, a Concentrated Feature Pyramid (CFP) is used to improve how the model fuses information from different scales, leading to a more thorough feature representation and thus more reliable detection. Our experiments show that the GAC-YOLO model performs exceptionally well, achieving 96.3% precision, 94.6% recall, and 95.2% mean Average Precision. This strong performance indicates our method is a practical and effective solution for insulator defect detection in real-world scenarios.