Improved Asphalt Pavement Crack Detection Model Based on Shuffle Attention and Feature Fusion

Abstract

Pavement distress is one of the most serious and prevalent diseases in pavement road detection. However, traditional methods for crack detection often suffer from low efficiency and limited accuracy, necessitating improvements in the accuracy of existing crack detection algorithms. Consequently, we propose the shuffle attention for you only look once version eight (SA-YOLOv8) model, which is based on an enhanced framework. Initially, we establish the required dataset and classify images proportionally based on their states. Subsequently, we conduct comparative testing against the results of the original model, analyzing issues such as the oversight of shallow and small cracks, truncation in the recognition of single-instance long cracks, and imprecise detection. We devise an improved detection approach based on YOLOv8. This method incorporates a small target detection layer to optimize the receptive field range, aiming to focus on identifying shallow and small cracks. Simultaneously, the Shuffle Attention mechanism and the transplanted spatial pyramid pooling-fast (SPP-F) reuse structure are introduced in the feature extraction network to enhance the model’s attention to detection targets. This augmentation improves the fusion of features for shallow small targets and overall and partial features of long cracks, thereby alleviating the precision of the model in crack detection. The experimental results demonstrate a stepwise improvement in the model’s mean average precision (mAP) with each enhancement to the original network. Initially, adding a small object detection layer increased the mAP by 3.4 percentage points, raising it to 68.2%. Subsequently, incorporating the spatial attention (SA) module resulted in a more substantial improvement, boosting the mAP by 8.7 percentage points to 73.5%. Finally, the addition of the transplanted SPP-F module further enhanced accuracy, increasing the mAP by 0.7 percentage points from the previous stage, thus achieving a final mAP of 74.2%. Overall, these modifications resulted in a total improvement of 9.4 percentage points in mAP compared to the original model. In conclusion, the proposed SA-YOLOv8s model effectively supports the automated recognition of asphalt road surface cracks, demonstrating applicability in practical scenarios. The recognition performance is notably favorable, demonstrating robustness in complex environments.