Answers are grounded in this paper and cite their sources.
YOLO-CSS improves defect detection accuracy and reduces computational costs in substation environments, highlighting its efficiency.
Detecting defects in substations is crucial for ensuring the reliability and safety of power systems. However, traditional detection methods often face challenges such as high computational costs, difficulties in handling complex environments with extreme scale variations among targets, densely clustered small components, and lowcontrast objects blending with backgrounds, particularly in substations exhibiting multiple types of defects. This paper proposes YOLO-CSS, which is a lightweight defect detection model that is specifically designed for complex substation scenarios. YOLO-CSS improves upon YOLOv8 by incorporating modules that strengthen feature extraction capabilities and improve robustness in complex environments by a more precise fusion of multiscale information. Additional modules were designed: one that adaptively focuses on critical feature regions in substation scenarios to enhance the defect detection capability and another that reduces the computational cost and memory usage through depthwise separable convolution to minimize the model size and facilitate real-time detection. A lightweight detection head is employed to improve defect detection accuracy and computational efficiency. Finally, Minimum Point Distance Intersection over Union is employed as the loss function to enhance the detection ability of small and dense objects. Experimental results demonstrated that YOLO-CSS reduced the number of parameters and computational cost by 32.8% and 27.2%, respectively, compared with YOLOv8 while demonstrating substantial improvements in the mean average precision at a model size of only 4.4MB. Furthermore, cross-dataset evaluations on industrial defect benchmarks validate the model’s superior generalization capability, confirming its effectiveness in diverse real-world scenarios beyond substation environments.
No takes yet. Share an insight, caveat, or question.
Ming et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: