EGA-YOLOv8n: An Improved YOLOv8 Approach for Drone-Based Wall Crack Detection

Detecting small objects in drone images is a difficult task because of their small size, the presence of intricate background and restricted amount of computational resources. Current detection methods have encountered challenges to trade-off between sensitivity and computational complexity. To address these problems, we propose EGA-YOLOv8n (ECA and GSConv Attention-YOLOv8n) in this work which is a small object detection algorithm based on the improved YOLOv8n model. First, GSConv is used to replace some conventional convolutions in the backbone network, which helps reduce parameters and computation. Meanwhile, an efficient lightweight attention mechanism ECA (Efficient Channel Attention) is introduced in the key layers of our model to enhance the ability of extracting slim small objects features such as cracks. Secondly, a Tiny Head branch is embedded into the detection head to perform direct detection with higher-resolution feature maps to improve the model’s perception towards extremely small objects. Empirical studies demonstrate that EGA-YOLOv8n can effectively improve the base performance of detecting small target cracks in drone aerial images significantly while maintaining its lightness, providing a feasible and efficient way for real-time deployment on drones.