Oil Leak Detection Method Based on Improved YOLOv7

In power plant environments, oil leaks can significantly impact production efficiency while posing serious safety hazards and environmental risks. Prompt and accurate detection of oil leaks is therefore a critical requirement in industrial operations. To overcome the limitations of traditional detection methods in handling complex backgrounds and small targets, this study presents an oil leak detection algorithm based on an enhanced YOLOv7 model. The original YOLOv7 algorithm struggles with adaptability in industrial scenarios. To address this, the proposed approach optimizes the network structure, enhancing both detection accuracy and robustness to meet real-world demands. A Squeeze-and-Excitation attention mechanism is integrated to improve focus on critical target features while reducing interference from complex backgrounds, thereby enhancing detection performance. Additionally, deformable convolution is incorporated to dynamically adjust convolution kernel shapes, allowing better adaptation to small and irregular targets and improving the model’s perception capabilities. The loss function is also modified to Focal-EIoU Loss, which combines the benefits of focal loss and extended intersection over union. This adjustment improves anchor box quality, boosts regression accuracy, and accelerates convergence. Experimental results confirm that the enhanced model achieves superior precision and reliability in oil leak detection tasks, demonstrating its effectiveness for industrial applications.