YOLOv8-LSD: Improved YOLOv8 Focused on Small Target Information Extraction for Road Damage Detection

Detection of road damage is critical for improving traffic safety and extending the life of roads. Despite the enhanced accuracy offered by deep learning models such as YOLOv8 in the field of road damage detection, challenges persist due to factors like susceptibility to weather conditions and pavement composition, and the difficulty in accurately identifying and pinpointing minor damages in complex scenarios. To tackle these issues, we introduce an enhanced detection algorithm, YOLOv8-LSD. This algorithm incorporates a deformable attention mechanism and a large separable kernel attention framework, which heighten focus on key regions within the input feature map, thereby facilitating precise damage detection. Additionally, it employs spatial and channel reconstruction in the detection head to optimize convolution, ensuring meticulous feature extraction and representation. Experimental findings demonstrate that our enhanced algorithm elevates detection accuracy by 1% over the original YOLOv8 model.